tests: benchdnn: support --mode-modifer=M for graph driver

[wzt1997]

General

The PR enables --mode-modifer=M for benchdnn graph driver:

1. Avoid creating reference primitives as well as the memories and customized data filling.
2. For customized operations, handle integral inputs specially for indexing values.
3. Otherwise, use fixed value( 0x3F3F3F3F (0.747059f) ) for data filling.

./tests/benchdnn/benchdnn --graph --mode=F --batch=test_graph_all
Output template: perf,%engine%,%prb%,%-time%,%0time%
perf,cpu,--mode=F --graph --case=op/f32/abs.json,0.0561523,0.139608
perf,cpu,--mode=F --graph --case=op/f32/abs_bwd.json,0.0541992,0.064628
perf,cpu,--mode=F --graph --case=op/f32/add.json,0.0571289,0.0702186
perf,cpu,--mode=F --graph --in-shapes=1:- --case=op/f32/add.json,0.0563965,0.0670133
perf,cpu,--mode=F --graph --case=op/f32/avgpool.json,4.81152,10.1638
perf,cpu,--mode=F --graph --op-attrs=0:auto_pad:SAME_UPPER*kernel:2x2 --case=op/f32/avgpool.json,4.58423,6.78208
//....

Sanitizer test:

        Start 201: test_benchdnn_modeR_graph_ci_cpu
199/220 Test #201: test_benchdnn_modeR_graph_ci_cpu ........................   Passed   21.13 sec

        Start 233: test_benchdnn_modeR_graph_ci_gpu
233/252 Test #233: test_benchdnn_modeR_graph_ci_gpu .........................   Passed   73.62 sec

[wzt1997]

make test
set test_scope=NIGHTLY
disable benchdnn_all
enable benchdnn_graph

[wzt1997]

make test
set test_scope=NIGHTLY
disable benchdnn_all
enable benchdnn_graph

[dzarukin]

It seems that internal mem_ can be constructed without lt_id_2_mems_.
I suggest to change the logic to following:

    // Emplace a memory object first.
    partition_mem_map.emplace(
                    id, dnn_graph_mem_t(lt_id_2_lt_.at(id), true));
    if (lt_id_2_mems_.at(id)) {
        SAFE(fill_underlying_mem_with_data(partition_mem_map[id], lt_id_2_mems_.at(id)), WARN);
    }

This way the flow is always the same and no_ref_memory will affect only the if part transparently without hurting a general logic.

[wzt1997]

For correctness check, we need to create graph memories based on primitive memory layouts, as benchdnn will reorder the outputs from both paths to abx during comparison. Therefore, the memory tags of outputs from two paths have to be identical.

For simplicity, I have done the following changes:

1. Move the data filling part to an independent function.
2. Remove the function that creates graph memories from graph logical tensors, and keep the primitive memory parameter.

[wzt1997]

make test
set test_scope=NIGHTLY
disable benchdnn_all
enable benchdnn_graph