[GPU][Loop] Change condition to reinterprete buffer (#25086)

### Details:
- In primitive_inst.cpp when skippable node which is optimized out, is
returned at realloc_if_needed(), output memory layout should be changed.
- So it needs to adjust in respective primitive_inst (for this case,
Reorder)

### Tickets:
 - 143848

Author: davidsnam-intel

src/plugins/intel_gpu/src/graph/reorder.cpp

@@ -270,7 +270,9 @@ void reorder_inst::update_output_memory() {
     if (!can_be_optimized())
         return;
 
-    if (static_cast<bool>(_outputs[0]) && _network.get_engine().is_the_same_buffer(output_memory(), input_memory()))
+    if (static_cast<bool>(_outputs[0])
+        && _network.get_engine().is_the_same_buffer(output_memory(), input_memory())
+        && output_memory().get_layout().identical(get_output_layout()))
         return;
 
     if (_node != nullptr)

src/plugins/intel_gpu/tests/unit/test_cases/loop_gpu_test.cpp

@@ -875,6 +875,115 @@ static void test_loop_gpu_multiple_shapes(ov::PartialShape body_input_layout,
     }
 }
 
+static void test_loop_gpu_multiple_shapes_single_shared(ov::PartialShape body_input_layout,
+                                        std::vector<ov::PartialShape> whole_layouts,
+                                        std::vector<std::vector<float>> input_data_list,
+                                        std::vector<float> expected_output_data,
+                                        int32_t axis,
+                                        size_t exit_value,
+                                        bool is_caching_test = false) {
+    auto& engine = get_test_engine();
+
+    auto b_input_layout = cldnn::layout{ body_input_layout, data_types::f32, format::bfyx };
+    auto const_layout = cldnn::layout{ {}, data_types::i64, format::bfyx };
+
+    auto e_initial_condition_mem = engine.allocate_memory(const_layout);
+    auto e_num_iteration_mem = engine.allocate_memory(const_layout);
+    auto b_exit_value_mem = engine.allocate_memory(const_layout);
+    auto b_index_inc_mem = engine.allocate_memory(const_layout);
+
+    // initialize input buffers
+    set_values(e_initial_condition_mem, {1});
+    set_values(b_exit_value_mem, {exit_value});
+    set_values(b_index_inc_mem, {1});
+    set_values(e_num_iteration_mem, {10});
+
+    primitive_id body_current_iteration_id = "b_index";
+    primitive_id body_execution_condition_id = "b_cond_exit_value";
+
+    cldnn::topology body(
+        input_layout(body_current_iteration_id, const_layout),
+        input_layout("b_parameter", b_input_layout),
+        data("b_exit_value", b_exit_value_mem),
+        data("b_index_inc", b_index_inc_mem),
+        eltwise("b_index_update", input_info(body_current_iteration_id), input_info("b_index_inc"), eltwise_mode::sum),
+        eltwise("b_permute", input_info("b_parameter"), input_info("b_index_update"), eltwise_mode::sum),
+        reorder("b_result", input_info("b_permute"), b_input_layout),
+        eltwise(body_execution_condition_id, input_info(body_current_iteration_id), input_info("b_exit_value"), eltwise_mode::lt)
+    );
+
+    primitive_id trip_count_id = "";
+    primitive_id actual_iteration_count_id = "actual_iteration_count";
+    primitive_id initial_condition_id = "initial_condition";
+    int64_t num_iterations = -1;
+
+    std::vector<loop::io_primitive_map> input_primitive_maps {
+        loop::io_primitive_map("input", "b_parameter", axis),
+        loop::io_primitive_map(actual_iteration_count_id, body_current_iteration_id) };
+    std::vector<loop::io_primitive_map> output_primitive_maps {
+        loop::io_primitive_map(cldnn::input_info("loop"), cldnn::input_info("b_result"), axis) };
+    std::vector<loop::backedge_mapping> back_edges {
+        loop::backedge_mapping("b_result", "b_parameter"),
+        loop::backedge_mapping("b_index_update", body_current_iteration_id) };
+
+    auto body_program = build_program(engine, body, body_execution_condition_id, output_primitive_maps, back_edges, true);
+
+    auto const_shape = engine.allocate_memory({ov::PartialShape{4}, data_types::i32, format::bfyx});
+    std::vector<int32_t> body_input_layouts;
+    for (size_t i = 0; i < body_input_layout.size(); i++) {
+        if (body_input_layout[i].is_dynamic())
+            body_input_layouts.push_back(-1);
+        else
+            body_input_layouts.push_back(body_input_layout[i].get_length());
+    }
+    set_values<int32_t>(const_shape, body_input_layouts);
+
+    cldnn::topology topology(
+        input_layout("input_origin", b_input_layout),
+        input_layout(initial_condition_id, e_initial_condition_mem->get_layout()),
+        mutable_data(actual_iteration_count_id, e_num_iteration_mem),
+        permute("input2", input_info("input_origin"), {0, 1, 2, 3}),
+        data("const", const_shape),
+        permute("permute1", input_info("input_origin"), {0, 1, 2, 3}),
+        concatenation("input", {input_info("permute1"), input_info("input_origin")}, 0),
+        loop("loop",
+             {input_info(actual_iteration_count_id), input_info(initial_condition_id), input_info("input")},
+             body_program, trip_count_id, initial_condition_id, actual_iteration_count_id,
+             input_primitive_maps, output_primitive_maps, back_edges,
+             num_iterations, body_current_iteration_id, body_execution_condition_id, 1),
+        permute("result", input_info("loop"), {0, 1, 2, 3}));
+
+    ExecutionConfig config = get_test_default_config(engine);
+    config.set_property(ov::intel_gpu::allow_new_shape_infer(true));
+
+    network network(engine, topology, config);
+    for (size_t i = 0 ; i < whole_layouts.size(); i++) {
+        auto whole_layout = whole_layouts[i];
+        auto input_data = input_data_list[i];
+
+        set_values(e_initial_condition_mem, {1});
+        set_values(b_exit_value_mem, {exit_value});
+        set_values(b_index_inc_mem, {1});
+        set_values(e_num_iteration_mem, {10});
+
+        auto e_input_layout = cldnn::layout{ whole_layout, data_types::f32, format::bfyx };
+        auto e_input_mem = engine.allocate_memory(e_input_layout); // b,f,x,y
+        auto expected_output_layout = whole_layout;
+        set_values(e_input_mem, input_data);
+
+        network.set_input_data("input_origin", e_input_mem);
+        network.set_input_data(initial_condition_id, e_initial_condition_mem);
+
+        auto outputs = network.execute();
+        auto output_layout = outputs.begin()->second.get_layout();
+        auto input_layout = network.get_primitive("input")->get_output_layout();
+
+        ASSERT_EQ(output_layout.feature(), input_layout.feature());
+        ASSERT_EQ(output_layout.spatial(0), input_layout.spatial(0));
+        ASSERT_EQ(output_layout.spatial(1), input_layout.spatial(1));
+    }
+}
+
 std::vector<float> input_data_2_4{
     1.0f,  2.0f,
     4.0f, -15.f,
@@ -919,6 +1028,15 @@ TEST(loop_gpu, support_loop_w_dynamic_input_w_various_shapes2) {
         -1, 10);
 }
 
+TEST(loop_gpu, support_loop_w_dynamic_input_w_various_shapes3) {
+    test_loop_gpu_multiple_shapes_single_shared(
+        { 1, -1, 560 },
+        {{ 1, 58, 560 }, { 1, 87, 560 }, { 1, 72, 560 }, { 1, 88, 560 }, { 1, 89, 560 }},
+        {input_data_2_4_4, input_data_2_4_4, input_data_2_4_4, input_data_2_4_4, input_data_2_4_4},
+        std::vector<float>(),
+        -1, 20);
+}
+
 static void test_loop_gpu_wo_trip_count_update_primitive_id(ov::PartialShape body_input_layout,
                                         std::vector<ov::PartialShape> whole_layouts,
                                         std::vector<std::vector<float>> input_data_list,