[SDPA] Guard mem efficient attention in deterministic mode (#91979)
# Summary
Memory efficient attention is a non deterministic algorithm.
This PR ensures that the sdp_choice will allow for mem-efficient to be used as the backend to SDPA if we are in warn only mode. Otherwise if we have enabled determinism and and set warn_only to False sdp_choice will not return memory efficient attention as the backend.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/91979
Approved by: https://github.com/cpuhrsch
