Derive probability of set_subtensor operations

[ricardoV94]

The logprob for the following kind of graphs can now be derived:

import pytensor.tensor as pt
import pymc as pm

x = pt.zeros((3, 3))
y = x[[0, 1, 2], [0, 1, 2]].set(pm.Normal.dist(shape=(3,)))
z = y[[0, 1, 2], [2, 1, 0]].set(pm.StudentT.dist(nu=2, shape=(3,)))

pm.logp(z, pt.zeros((3, 3))).eval().round(2)
# array([[-0.92,  0.  , -1.04],
#        [ 0.  , -1.04,  0.  ],
#        [-1.04,  0.  , -0.92]])

[codecov]

Codecov Report

Attention: Patch coverage is 94.94949% with 5 lines in your changes missing coverage. Please review.

Project coverage is 92.87%. Comparing base (3057351) to head (75cf84d).
Report is 45 commits behind head on main.

Files with missing lines	Patch %	Lines
pymc/logprob/set_subtensor.py	94.11%	5 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #7553      +/-   ##
==========================================
+ Coverage   92.83%   92.87%   +0.04%     
==========================================
  Files         106      107       +1     
  Lines       17669    17749      +80     
==========================================
+ Hits        16403    16485      +82     
+ Misses       1266     1264       -2     

Files with missing lines	Coverage Δ
pymc/logprob/__init__.py	100.00% <100.00%> (ø)
pymc/logprob/checks.py	100.00% <100.00%> (ø)
pymc/logprob/utils.py	97.14% <100.00%> (+4.67%)	⬆️
pymc/logprob/set_subtensor.py	94.11% <94.11%> (ø)

[ricardoV94]

Should restrict to when updates do not depend on previous writes to the buffer. Otherwise logp is not as straightforward/always deriverable