Don't try to infer support shape of multivariate RVs by default

[ricardoV94]

The default method is not robust and can lead to cryptic error messages.

In a discussion with a user https://discourse.pymc.io/t/pm-sample-prior-predictive-fails-with-incorrect-size-info-passed-to-the-rng-fn-function-of-the-randomvariable/12513 , they were trying to fake a multivariate distribution as a univariate because of the cryptic error induced by the default method.

We now raise NotImplementedError by default, and include illustrative examples in the docstrings that should help users implementing this method.

Also fixed a couple of minor bugs.

The changes will break RVs that relied on the default implementation. This should be easy to fix and won't go unnoticed.

[codecov-commenter]

Codecov Report

Merging #388 (c689549) into main (9be43d0) will decrease coverage by 0.01%.
The diff coverage is 97.82%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #388      +/-   ##
==========================================
- Coverage   80.46%   80.46%   -0.01%     
==========================================
  Files         156      156              
  Lines       45514    45517       +3     
  Branches    11148    11148              
==========================================
+ Hits        36624    36626       +2     
- Misses       6688     6689       +1     
  Partials     2202     2202              

Impacted Files	Coverage Δ
pytensor/tensor/random/op.py	96.62% <93.75%> (-0.87%)	⬇️
pytensor/tensor/random/basic.py	99.04% <100.00%> (+<0.01%)	⬆️
pytensor/tensor/random/utils.py	100.00% <100.00%> (ø)

[michaelosthege]

LGTM