Stack array stats if possible

[aloctavodia]

BART (that now lives in pymc-experimental) stores 2 stats which at each point are arrays. Variable importance and bart_trees. For example variable importance is a vector with the position encoding a given covariate and an integer encoding the number of times that covariate was used in the trees.

We used to have this to stack those stats https://github.com/pymc-devs/pymc/pull/5566/files#diff-7b3470589153c7c90a2c05fc86402f9acf65555200d90f01b51a47ccef592628L637-L654 but that code is BART-specific, so here I am proposing as alternative a more general solution.

[codecov]

Codecov Report

Merging #5572 (84d9f8f) into main (44c5495) will not change coverage.
The diff coverage is 100.00%.

❗ Current head 84d9f8f differs from pull request most recent head 3f922a3. Consider uploading reports for the commit 3f922a3 to get more accurate results

@@           Coverage Diff           @@
##             main    #5572   +/-   ##
=======================================
  Coverage   87.59%   87.59%           
=======================================
  Files          76       76           
  Lines       13694    13694           
=======================================
  Hits        11995    11995           
  Misses       1699     1699           

Impacted Files	Coverage Δ
pymc/backends/base.py	87.77% <100.00%> (+0.09%)	⬆️
pymc/step_methods/mlda.py	96.37% <100.00%> (-0.03%)	⬇️

[OriolAbril]

Code looks good but I don't really know how it will integrate with the rest of the library, hope these two questions help a bit.

Do we have tests with compound steps similar to #4602, does this modify this?

Would it be helpful to try and provide stepper defined dims or coords?

[aloctavodia]

This is not directly related to compound steps, but to samplers/steps that have stats that are arrays and not floats/integer/boolean etc. So far it seems that the only samplers with that kind of stats are from PGBART and MLDA.

[michaelosthege]

Would it be helpful to try and provide stepper defined dims or coords?

Generally yes. Alongside shape information.

@aloctavodia why do you want to stack these variables at all?
They should be yielded in the correct type right away. We don't have a stats-postprocessing API and I'd prefer to undertstand what you're trying to achieve first.

[michaelosthege]

I didn't check why the CI failed, but I'm quite certain about the isinstance checks I commented above.