Fix bug from too strict type check in _sum_grad_over_bcasted_dims

[ricardoV94]

Fix bug in gradient of IncSubtensor reported in https://discourse.pymc.io/t/marginalized-mixture-wont-begin-sampling-throws-assertion-error/15969/10?u=ricardov94

---

📚 Documentation preview 📚: https://pytensor--1036.org.readthedocs.build/en/1036/

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 81.93%. Comparing base (4c7b494) to head (636cdca).
Report is 97 commits behind head on main.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #1036   +/-   ##
=======================================
  Coverage   81.93%   81.93%           
=======================================
  Files         182      182           
  Lines       47890    47889    -1     
  Branches     8617     8617           
=======================================
+ Hits        39239    39240    +1     
+ Misses       6481     6480    -1     
+ Partials     2170     2169    -1     

Files with missing lines	Coverage Δ
pytensor/tensor/subtensor.py	89.48% <100.00%> (+0.17%)	⬆️

[lucianopaz]

In all honesty, I don’t understand what was wrong with the past asserts and why the new conditionals are better, but the test you added looks convincing so I’ll approve

[ricardoV94]

The old assert assumed you had the same knowledge about broadcastability of the input and the gradient wrt to the input.

The new ones assert you have the same or more (but not less, because if you knew the input was broadcastable, you should also know the gradient will be)

Broadcastable just means static shape = 1 along that dim.