Consider to combine enable_fp_fusion and enable_fast_math in upstream triton

[minjang]

#41 can be upstreamed? Do more thorough testing for GPU and try to make a PR.

[embg]

One data point: in clang, -ffast-math does enable fp fusion: https://clang.llvm.org/docs/UsersManual.html#cmdoption-ffast-math

This option implies:
...

• -ffp-contract=fast

Perhaps we can have three settings for enable_fp_fusion? If "enabled" or "disabled", we don't touch it. But if it is set to "auto", then we enable or disable depending on whether fast-math is enabled.