Investigate whether pad mm is useful on XPU

[jianyizh]

inductor may pad mm for some shapes i.e replace matmul with zeros, cat, matmul and slice. On A100 fp16 fp_GPT2, this feature can improve 30% performance. Check if it's useful on XPU. https://github.com/pytorch/pytorch/blob/main/torch/_inductor/fx_passes/pad_mm.py

[jianyizh]

General guideline from onednn team: pad to a multiple of 64 bytes (1 cache line) in the unit-stride dimension, but try to avoid multiples of large powers of 2 (say 2048 bytes).

So for bf16, 37 elements in contiguous dimension = 74 bytes -> pad to 128 bytes, but 1023 elements = 2046 bytes -> pad to 2048 + 64 bytes to avoid a multiple of 2048.

It's enough to pad the tensor strides; you do not need to pad the tensor size itself if you don't want to.