transformations: (lift-arith-to-linalg) Add generic FMA

[n-io]

Adds the option to lift arith to generic fused multiply-add ops, with various flags to control behaviour.

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 90.08%. Comparing base (a3e2e7c) to head (542da59).
Report is 3 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #3344   +/-   ##
=======================================
  Coverage   90.08%   90.08%           
=======================================
  Files         446      446           
  Lines       56419    56457   +38     
  Branches     5417     5421    +4     
=======================================
+ Hits        50824    50862   +38     
  Misses       4162     4162           
  Partials     1433     1433           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[superlopuh]

This feels like it's doing two things at the same time, and breaking the composability of rewrites. Could the ops be fused by a separate pass?

[n-io]

The behaviour can be switched on and off by a flag, but doing this in a separate pass is (mildly) more complicated, as the regions contain arith.addf and arith.mulf which the pass as-is would attempt to translate.

[AntonLydike]

I agree with Sasha here, lifting arith to linalg is one thing, but fusing operations together is a different thing entirely

[n-io]

Closing in favour of doing this in a separate pass.