vulkan: optimized coopmat matmul perf for IntelGPU

[fish-jiang]

Summary
In this PR we

1. enabled coopmat on Xe1
2. did warp tuning on Xe2/Xe3
3. optimized coopmat1 shader code to reduce unnecessary memory load and tensor core calculation for MoE model.

[h9j6k]

Bringing older xe devices in line is great. Does the 'Xe1' you are referring to include intel DG1, the very first gen of discrete intel xe? Or this Xe1 coopmat enablement is for DG2 onwards like alchemist A770/A380 etc (gfx 12.5+)? Thanks.