MHA optimizations (#93234)
Slight perf optimizations for regular MHA by reducing the number of kernels called
Before:
After:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/93234
Approved by: https://github.com/drisspg
