Enforce dims elements to be strings

[ricardoV94]

I think there was a confusion introduced in 64e63d8

Unless I am missing something, it's not possible to have arbitrary missing dims. Only works if missing dims are at the rightmost positions.

Closes #6379

[codecov]

Codecov Report

Merging #6470 (41944f7) into main (00d2675) will increase coverage by 0.00%.
The diff coverage is 100.00%.

Additional details and impacted files

@@           Coverage Diff           @@
##             main    #6470   +/-   ##
=======================================
  Coverage   94.79%   94.79%           
=======================================
  Files         148      148           
  Lines       27730    27732    +2     
=======================================
+ Hits        26287    26289    +2     
  Misses       1443     1443           

Impacted Files	Coverage Δ
pymc/backends/arviz.py	96.37% <100.00%> (ø)
pymc/model.py	89.86% <100.00%> (+0.02%)	⬆️

[michaelosthege]

I think #6379 was accurate: The problem was only in the InferenceData conversion and the solution Oriol suggested here should fix it.
The test you wrote should be able to confirm that

IIRC there are some valid use cases for None dims with multivariate RVs such as a MultivariateNormal, that's why we (and ArviZ) had this in the first place

[ricardoV94]

There are other issues with None. For instance, model.add_coord assumes all dims are a string, but we are calling it with None dname from model.register_rv. It seems like the infrastructure to allow unspecified dims was not really worked out completely.

[OriolAbril]

looks good.

In case it helps, here is some extra context on arviz and xarray. Neither of them support None as dimensions, but they support in some cases not providing dimensions or providing only part of them (in which case, the provided dims will be used to name the first n axis, the n+1 and more axis get default dim names). Moreover, xarray allows any hashable to be used as dimension name, not necessarly strings. ArviZ aims to not break when dims aren't strings (using reprs and str fallbacks if necessary) but it kind of assumes that dimension names are strings.