Fix rejection-based truncation of scalar variables

[ricardoV94]

Reported in pymc-devs/pytensor#442

---

📚 Documentation preview 📚: https://pymc--6923.org.readthedocs.build/en/6923/

[codecov]

Codecov Report

Merging #6923 (5fa5721) into main (df7b267) will increase coverage by 0.38%.
The diff coverage is 100.00%.

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #6923      +/-   ##
==========================================
+ Coverage   91.78%   92.17%   +0.38%     
==========================================
  Files         100      100              
  Lines       16845    16847       +2     
==========================================
+ Hits        15462    15528      +66     
+ Misses       1383     1319      -64     

Files	Coverage Δ
pymc/distributions/truncated.py	99.41% <100.00%> (+<0.01%)	⬆️

... and 4 files with indirect coverage changes

[lucianopaz]

This looks very nice @ricardoV94. I might need you to guide me a bit through the logic here because I haven't grasped the subtleties of trying to truncate SymbolicDistributions yet.

[lucianopaz]

Shouldn't this also have some kind of pt.and_(not(reject_draws), ... so that the draws that were already accepted don't get resampled? I fail to see where that is happening.

[ricardoV94]

The draws that were already accepted will always be within upper and lower. The set_subtensor only changes the indexes that were not already valid.

[lucianopaz]

Will this until cut the scan short of the max_n_steps if the condition is met sooner?

[ricardoV94]

Yes, that's the whole point of the until (Also you can never define one without max_n_steps if that helps)

[ricardoV94]

@lucianopaz, I realized I needed a bigger refactor, mostly because pymc-devs/pytensor#473 makes it hard to box other SymbolicRVs safely.

This PR is now just fixing the bug with the scalar case, and I'll open another one later with then new functionality