Enable no-cpython-wrapper in numba where possible

[aseyboldt]

Disabling the generation of cpython wrappers should speed up compilation quite a bit.
Most of our intermediate functions are never called from python, so we don't need numba to generate wrapper code.

Note, that calling a njit function with no_cpython_wrapper directly from python (so not within a different njit function) will not work, and usually segfault.

The compilation time should further improve, if something like numba/numba#9566 lands in numba.

[ricardoV94]

Sweet

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 80.83%. Comparing base (db73461) to head (945dc09).
Report is 1 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main     #765   +/-   ##
=======================================
  Coverage   80.83%   80.83%           
=======================================
  Files         162      162           
  Lines       46860    46864    +4     
  Branches    11465    11465           
=======================================
+ Hits        37877    37883    +6     
+ Misses       6732     6730    -2     
  Partials     2251     2251           

Files	Coverage Δ
pytensor/link/numba/dispatch/basic.py	85.74% <100.00%> (+0.06%)	⬆️
pytensor/link/numba/dispatch/elemwise.py	88.72% <100.00%> (ø)
pytensor/link/numba/linker.py	100.00% <100.00%> (ø)

... and 1 file with indirect coverage changes

[ricardoV94]

How much speedup did you observe with this?

[aseyboldt]

Compilation took a bit more than half as long with the change applied for two models I tried. Would be curious what other models do though.

[ricardoV94]

Compilation took a bit more than half as long with the change applied
for two models I tried. Would be curious what other models do though.

And do you expect further speed up if the numba changes get merged or this achieves mostly the same?

[aseyboldt]

And do you expect further speed up if the numba changes get merged or this achieves mostly the same?

The model I've been using as a baseline went from 20s to 13s with this PR, and then to 8s with the numba PR. But I don't know if that is representative. (All measured hopefully without any caches).

(somehow I managed to edit a comment instead of replying, sorry)