[SYCL][NFCI] Finalize switch to SPV_KHR_cooperative_matrix

[MrSidims]

No description provided.

[MrSidims]

Tested locally on linux both CPU and GPU - the patch shouldn't cause any regressions (unless there is a difference between driver versions used in pre-commit here in CI and drivers used locally (setup by scripts)).
Hasn't yet tested on windows - having troubles finding DG2 machine.

[YuriPlyakhin]

Code changes itself LGTM.

@MrSidims , in addition to stability, could you please also verify performance just in case to make sure no performance regressions are introduced?

[dkhaldi]

Let's investigate failures on Windows DG2 before merging this

[MrSidims]

@dkhaldi on windows DG2 only these 2 tests are failing:
Failed Tests (2):
SYCL :: Matrix/SPVCooperativeMatrix/joint_matrix_bf16_fill_k_cache_SLM.cpp
SYCL :: Matrix/joint_matrix_bf16_fill_k_cache_SLM.cpp

But there are indeed regressions. joint_matrix_out_bounds.cpp was added only for JointMatrices, with the switch it's actually failing. The reason is that __spirv_CooperativeMatrixLoadCheckedINTEL (and Store) are not handled by IGC for Cooperative Matrices at all. This behavior doesn't follow SPIR-V specification as it defines OpCooperativeMatrixLoadCheckedINTEL be used only for OpTypeCooperativeMatrixKHR. Support of TypeJointMatrixINTEL was left in the translator only as an exception for the switch. Prefetch seem to also not handled by IGC when used with TypeCooperativeMatrixKHR. We have to postpone the switch to cooperative matrices as well as promoting the sycl extension from experimental to supported. Note, only new functionality doesn't work. Old functionality works as expected.

[MrSidims]

Closing the PR as we can't merge it