Updated moe to work with different models

[teubert]

Previously MoE only worked with homogenous models. Now if a output, etc. is missing it will keep trying the next best model until all are completed.

[github-actions]

Thank you for opening this PR. Each PR into dev requires a code review. For the code review, look at the following:

• Reviewer (someone other than author) should look for bugs, efficiency, readability, testing, and coverage in examples (if relevant).
• Ensure that each PR adding a new feature should include a test verifying that feature.
• All errors from static analysis must be resolved.
• Review the test coverage reports (if there is a change) - will be added as comment on PR if there is a change
• Review the software benchmarking results (if there is a change) - will be added as comment on PR
• Any added dependencies are included in requirements.txt, setup.py, and dev_guide.rst (this document)
• All warnings from static analysis must be reviewed and resolved - if deemed appropriate.

[codecov-commenter]

Codecov Report

Merging #110 (0ac867c) into dev (70ab878) will not change coverage.
Report is 5 commits behind head on dev.
The diff coverage is 0.00%.

❗ Current head 0ac867c differs from pull request most recent head fb3f5d2. Consider uploading reports for the commit fb3f5d2 to get more accurate results

@@           Coverage Diff           @@
##              dev     #110   +/-   ##
=======================================
  Coverage   83.54%   83.54%           
=======================================
  Files         100      100           
  Lines       10281    10281           
=======================================
  Hits         8589     8589           
  Misses       1692     1692           

Files	Coverage Δ
src/progpy/mixture_of_experts.py	13.54% <0.00%> (ø)

[matteocorbetta]

Looks good and works as expected. it took me a little to understand what was going on tho.
Can the idea behind the heterogeneous comparison with multi-output models be clarified in the test_moe with more comments? Maybe I did not understand it. If models have different output size, the only thing you can do is rank their performance per output, is that what's happening?
I think giving a practical example, instead of random numbers (x0+b, x0+c, with b=0.75, etc) would help the user understand what this is for.