apply simple sdpa to coreml/mps backend (#3660)

Summary:
Pull Request resolved: https://github.com/pytorch/executorch/pull/3660

coreml and mps doesn't support sdpa at the moment, use simple sdpa to have a simpler decomposition. Observer 1.5x faster on emulator.

Reviewed By: shoumikhin, kirklandsign

Differential Revision: D57476985

fbshipit-source-id: 2dbcad1a6e8b744e0a95d60fcd740369d665eab0