change training precision default to tf32
Summary: PyTorch changed tf32 off by default but for internal model training we use tf32
Reviewed By: xuzhao9, brad-mengchi
Differential Revision: D40008269
fbshipit-source-id: f6a5a099245f874c7c73ed009ed40256d6dbfbf5
