xe: sdpa: add support for reusable sdpa

[syurkevi]

Description

This PR introduces support for reusable sdpa so recompilation can be skipped for different sequence and query lengths. Head size is still baked into the kernel. The configuration for microkernel headers has been extracted and made part of the headers since this must now be done ahead of execution for caching.

Checklist

General

• Do all unit and benchdnn tests (make test and make test_benchdnn_*) pass locally for each commit?
• Have you formatted the code using clang-format?

[syurkevi]

<3% geomean regression from reusable params in kernel.
Example of reusable cache hit:

onednn_verbose,v1,primitive,create:cache_miss,gpu,sdpa,ocl:micro:any,undef,query:f16::blocked:abcd::f0 key:f16::blocked:abdc::f0 val:f16::blocked:abcd::f0 msk:f16::blocked:abcd::f0 dst:f16::blocked:abcd::f0,,msk:1d,1x4x1x128:1x4x128x386:1x4x386x128,6.40015

onednn_verbose,v1,primitive,create:kernel_cache_hit,gpu,sdpa,ocl:micro:any,undef,query:f16::blocked:abcd::f0 key:f16::blocked:abdc::f0 val:f16::blocked:abcd::f0 msk:f16::blocked:abcd::f0 dst:f16::blocked:abcd::f0,,msk:1d,1x4x1x128:1x4x128x391:1x4x391x128,0.00683594

[syurkevi]

make test
disable benchdnn_all
enable benchdnn_graph
enable test_device_gpu
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg

[syurkevi]

make test
disable benchdnn_all
enable benchdnn_graph
enable test_device_gpu
enable arch_gpu_xe-hpc
enable arch_gpu_xe-hpg-atsm
enable arch_gpu_xe-hpg-dg2
enable arch_gpu_xe-lpg+
enable arch_gpu_xe2-hpg-bmg
enable arch_gpu_xe2-lpg