Remove support for continuation of traces

[michaelosthege]

Choosing initial values for MCMC chains gets a lot easier if we stop supporting to start from existing traces.

It will also make it easier to swap backends in the future, when the public API no longer accepts MultiTrace objects.

There are more good reasons why continuation of traces is bad:

• Properties like sampling time no longer apply
• Model could have changed inbetween
• Tuning will be different in the first and second part of the final result

Starting from existing posterior samples can still be achieved via the pm.sample(start=...) kwarg.

Concatenation of sampling results should be done in a postprocessing step.

Depending on what your PR does, here are a few things you might want to address in the description:

• what are the (breaking) changes that this PR makes? 👉 Removes support for a little known feature that's not a good practice to begin with.
• important background, or details about the implementation 👉 Docstrings and type hints were edited. Errors are raised in _choose_backend which is always involved.
• are the changes—especially new features—covered by tests and docstrings? 👉 Yes.
• linting/style checks have been run
• consider adding/updating relevant example notebooks
• right before it's ready to merge, mention the PR in the RELEASE-NOTES.md Will add this in the release notes draft document manually.

[codecov]

Codecov Report

Merging #5019 (b2b21c6) into main (53e572c) will decrease coverage by 2.11%.
The diff coverage is 90.00%.

@@            Coverage Diff             @@
##             main    #5019      +/-   ##
==========================================
- Coverage   77.24%   75.13%   -2.12%     
==========================================
  Files          85       85              
  Lines       14004    14004              
==========================================
- Hits        10818    10522     -296     
- Misses       3186     3482     +296     

Impacted Files	Coverage Δ
pymc3/sampling.py	87.01% <90.00%> (+0.12%)	⬆️
pymc3/step_methods/mlda.py	11.85% <0.00%> (-74.49%)	⬇️
pymc3/tests/models.py	82.01% <0.00%> (-5.76%)	⬇️