Add triton intel gpu gather op lowering.

[chengjunlu]

Add triton intel gpu gather op lowering.

[chengjunlu]

Hi @sarah12121212 ,

This is just for your information:
This is the code lowering we discussed about the optimization for the fused moe.

The fast path is using the inline VISA to embad the gather IO to load the A matrix as DPAS layout from the memory.
The default path is to use the LLVM load operation which is interpreted as gather IO of IGC backend.

I need to switch to other tasks. If it blocked the vLLM tasks, please let me know.