Minor optimizations and improvements

[bvdmitri]

This PR

• Removes a special structure for UnspecifiedConstraints and simply uses Constraints (UnspecifiedConstraints is a const now)
• Removes a special structure for BetheFactorization and simply uses UnspecifiedConstraints
• Add a clarification to MeanField, that it does not respect default_constraint
• Improves the allocation profile in some places by doing more operations in place rather than allocating a new BitSet matrices (checked on a simple model on RxInfer)
• Moved "Apply MeanField constraints" and "Apply BetheFactorization constraints" to a separate file

Note: Benchmarks show that getorcreate! is significantly slower (and only in case of 1), I don't believe it and it seems like a noise. I didn't touch this part of the code, so it should not become slower or faster.

Would be nice if you could double check it @wouterwln

[bvdmitri]

This actually a funny one @wouterwln , not sure who of us made this mistake, but it basically created a vector of vector of constraints instead of just vector of constraints. It somehow worked though (probs some weird convert mechanics)

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 91.74%. Comparing base (918bbbc) to head (4bc587a).
Report is 7 commits behind head on dev-4.0.0.

Additional details and impacted files

@@              Coverage Diff              @@
##           dev-4.0.0     #173      +/-   ##
=============================================
+ Coverage      91.69%   91.74%   +0.04%     
=============================================
  Files             14       14              
  Lines           1795     1781      -14     
=============================================
- Hits            1646     1634      -12     
+ Misses           149      147       -2     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[wouterwln]

On my machine getorcreate was also slower, but this was already before my implementation of MeanField. I thought it was noise as well so I merged, but there might be more stuff going on? In any case. It shouldn't matter as long as pipeline benchmarks aren't affected