Faster infer_static_shape

[ricardoV94]

This method which is called every time we create an Alloc or RandomVariable Op, can be pretty slow, because it calls canonicalize + constant_folding on the shape graph, which can be pretty large sometimes. This can also be pretty wasteful if the shapes are already constant to begin with!

The utility is now much more targeted, using only rewrites that are directly relevant for static shape propagation. It also avoids doing costly C-compilation

This should result in much snappier creation of RandomVariables and rewrites

[codecov-commenter]

Codecov Report

Merging #521 (77e597d) into main (7ecb9f8) will increase coverage by 0.01%.
The diff coverage is 91.78%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #521      +/-   ##
==========================================
+ Coverage   80.82%   80.84%   +0.01%     
==========================================
  Files         162      162              
  Lines       46194    46249      +55     
  Branches    11288    11318      +30     
==========================================
+ Hits        37337    37388      +51     
- Misses       6630     6633       +3     
- Partials     2227     2228       +1     

Files	Coverage Δ
pytensor/tensor/rewriting/basic.py	94.09% <100.00%> (+0.08%)	⬆️
pytensor/tensor/rewriting/math.py	86.50% <100.00%> (+<0.01%)	⬆️
pytensor/tensor/rewriting/shape.py	81.43% <100.00%> (+0.41%)	⬆️
pytensor/tensor/rewriting/subtensor.py	88.80% <100.00%> (+0.07%)	⬆️
pytensor/tensor/shape.py	93.09% <86.66%> (+0.01%)	⬆️
pytensor/tensor/basic.py	88.39% <87.50%> (-0.05%)	⬇️