sycl : variable sg_size support for mmvq kernels

[Alcpz]

MMVQ kernels were based on CUDA's mmvq kernels, which were tailored for subgroups of size 32. While bigger subgroups were considered, smaller subgroups were not. The changes allow for more flexibility in these kernels and reduce the code base size if we make these kernels fast enough in intel architectures.

The introduced changes distribute more work per thread / workitem if WARP_SIZE < 32.

This PR doesn't affect performance.

QK_WARP_SIZE is still needed for some kernels that still require subgroup size == 32, so I am renaming the existing variable for now.

Note: To enable these changes on Intel devices, further changes are required in the entry point of the MUL_MAT op.

Note2: This code path requires FMA intrinsics that may not be available for all architectures, even more changes are required to make it the default path on Intel devices without potentially losing support on older archs

[Rbiessy]

LGTM

[qnixsynapse]

Awesome. Thank you!
LGTM