Set FastMathFlags by default

[whitneywhtsang]

Move setting of fast math flags from ScheduleLoad pass, so that it can be reused for both default and advanced path.
CI: https://github.com/intel/intel-xpu-backend-for-triton/actions/runs/11302770929

[chengjunlu]

Should be the fast math controlled by the user for tuning the accuracy?

[whitneywhtsang]

Should be the fast math controlled by the user for tuning the accuracy?

According to the logic in triton-lang/triton@84fe9da, users are expected to use libdevice if they want high precision.

[chengjunlu]

Should be the fast math controlled by the user for tuning the accuracy?

According to the logic in triton-lang/triton@84fe9da, users are expected to use libdevice if they want high precision.

Make sense. Another question not related to this PR. It seems the default path has already uses the fast math because we use the same lowering logic as the upstream. Is it?

[whitneywhtsang]

Make sense. Another question not related to this PR. It seems the default path has already uses the fast math because we use the same lowering logic as the upstream. Is it?

Default path uses the same lowering logic as upstream, but it doesn't add fast math as aggressive as this PR. This PR adds to add possible operations that have the fast math interface.

[FMarno]

It would be nice to see something to evidence that this change makes a difference (like some tests) but the changes look fine.