graph: introduce internal dnnl_sdpa op

[xiang1guo]

Description

• This is the first part to support GQA pattern refinement requested by PyTorch upstream. Address MFDNN-12871. The PR added a new internal dnnl_sdpa op for better fusion and alignment with sdpa primitive. With this, the internal compilation will transform a subgraph into sdpa op, see following pic.
• The refactor also aims to reduce graph compilation time by reducing layout propagation and memory planning time with a simplified internal sdpa op.

Works

• Added a dnnl_sdpa op. Currently only support float SDPA.
• Added a new sdp_primitive_v1 kernel for simplicity. The final goal is to merge this kernel with sdp_primitive kernel.
• Move the sdpa primitive ukernel creation process into op_executable which is now aligned with other kernels

Follow-up

There will be another PR to refine the GQA pattern based on this new internal dnnl_sdpa.

validation

There are total 66 case can be supported by GPU ukernel, those 42 float SDPA cases can run into sdp_primitive_v1_kernel_t now, the other 24 cases are quantization SDPA which all run into sdp_primitive_kernel_t .

Kernel	PR	main branch
sdp_primitive_v1_kernel_t	42	0
sdp_primitive_kernel_t	24	66
larger_partition_kernel_t	76	76

Performance test is WIP.