Use more conservative fastmath flags in numba backend

[ricardoV94]

Closes #1145

---

📚 Documentation preview 📚: https://pytensor--1147.org.readthedocs.build/en/1147/

[codecov]

Codecov Report

Attention: Patch coverage is 83.33333% with 4 lines in your changes missing coverage. Please review.

Project coverage is 82.15%. Comparing base (8cc489b) to head (769d134).
Report is 2 commits behind head on main.

Files with missing lines	Patch %	Lines
pytensor/link/numba/dispatch/basic.py	57.14%	1 Missing and 2 partials ⚠️
pytensor/link/numba/dispatch/elemwise.py	80.00%	1 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1147      +/-   ##
==========================================
+ Coverage   82.14%   82.15%   +0.01%     
==========================================
  Files         186      186              
  Lines       48207    48193      -14     
  Branches     8677     8677              
==========================================
- Hits        39598    39593       -5     
+ Misses       6441     6430      -11     
- Partials     2168     2170       +2     

Files with missing lines	Coverage Δ
pytensor/link/numba/dispatch/blockwise.py	90.00% <ø> (ø)
pytensor/link/numba/dispatch/extra_ops.py	95.49% <100.00%> (-0.03%)	⬇️
pytensor/link/numba/dispatch/scalar.py	94.40% <100.00%> (-0.04%)	⬇️
pytensor/link/numba/dispatch/elemwise.py	94.79% <80.00%> (+2.80%)	⬆️
pytensor/link/numba/dispatch/basic.py	81.62% <57.14%> (-0.61%)	⬇️

... and 1 file with indirect coverage changes

[aseyboldt]

Maybe we should only use these (or some other subset) by default?

pytensor/pytensor/link/numba/dispatch/vectorize_codegen.py

Line 62 in 8cc489b

"fastmath": {

All fastmath flags is really aggressive, and leads to some undefined behavior like in the nan case...

[ricardoV94]

@aseyboldt I moved the fastmath subflag definition to the numba_njit wrapper

[aseyboldt]

Looks good :-)