[CUBLAS][TF32] Add environment variable to allow override of `allow_tf32_cublas` (#77114)
#76509 changes the default behavior of matmuls to avoid using TF32. However, TF32 use cases still exist including CI/deployment environments that are often managed via environment variables. This PR just adds an environment variable check, currently called `TORCH_ALLOW_TF32_OVERRIDE` to support enabling TF32 via an environment variable rather than the C++/Python API.
CC @xwang233 @ptrblck @syed-ahmed @csarofeen @ngimel @mruberry
Pull Request resolved: https://github.com/pytorch/pytorch/pull/77114
Approved by: https://github.com/ngimel