[ADD] Post-Hoc ensemble fitting

[ravinkohli]

Adds a fit ensemble method enabling users to post hoc build an ensemble after search is done. This feature allows the user to run a search and sequentially build an ensemble later. This PR also adds tests for the functionality as well as updating traditional_run_history.json.

[codecov]

Codecov Report

Merging #260 (4d55a7d) into development (8237f2c) will increase coverage by 0.15%.
The diff coverage is 78.43%.

@@               Coverage Diff               @@
##           development     #260      +/-   ##
===============================================
+ Coverage        81.76%   81.91%   +0.15%     
===============================================
  Files              151      151              
  Lines             8641     8677      +36     
  Branches          1326     1333       +7     
===============================================
+ Hits              7065     7108      +43     
+ Misses            1104     1096       -8     
- Partials           472      473       +1     

Impacted Files	Coverage Δ
autoPyTorch/api/base_task.py	84.02% <78.00%> (-0.63%)	⬇️
autoPyTorch/datasets/base_dataset.py	80.48% <100.00%> (+0.48%)	⬆️
...ipeline/components/setup/network_backbone/utils.py	87.21% <0.00%> (-1.51%)	⬇️
autoPyTorch/ensemble/ensemble_builder.py	73.58% <0.00%> (-0.63%)	⬇️
autoPyTorch/utils/stopwatch.py	70.10% <0.00%> (+1.03%)	⬆️
autoPyTorch/ensemble/ensemble_selection.py	96.84% <0.00%> (+1.05%)	⬆️
...Torch/pipeline/components/training/metrics/base.py	68.65% <0.00%> (+2.98%)	⬆️
autoPyTorch/optimizer/smbo.py	83.19% <0.00%> (+3.36%)	⬆️
...reprocessing/feature_preprocessing/TruncatedSVD.py	93.10% <0.00%> (+3.44%)	⬆️
... and 3 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 8237f2c...4d55a7d. Read the comment docs.

[nabenabe0928]

Thanks for the PR.
I left some comments.

[franchuterivera]

Thanks a lot for the changes. They look good but I am requesting mainly two things to make it even better:

Add an example where we fit the API sequentially. First search is called with 1 core, then fit the ensemble

Then add some extra checks as mentioned to gain more confidence on the chance.

[nabenabe0928]

I did not find large issues except what Francisco already mentioned.
I will wait for another major change.