Multistep search

[mschwoer]

version 1:

A first very rough outline, just to get some early feedback.

The idea is to activate the MSS via the config parameter
multistep_search: enabled: True
The same config key holds the special parameters required for each step, which overwrite default.yaml.
Orchestration of the workflow (and the folders) is done as code.

See individual commit messages for an outline of the ideas.

update 1:

Refactored according to feedback.
Decided still against having everything in one folder, as it makes life much harder: the prefix would need to be made > known to a lot of parts in the code, and if you miss one you might tacitly overwrite files (but open for discussion). The last > step will always write to "output_dir", such that the "final" results are always at the same place.

I am happy enough with the overall structure, among the open questions for this PR would be the exact settings in >multistep.yaml (see also the added TODOs which show some additional uncertainties).

update 2:
This PR adds the multistep search including tests, UI, docs, and e2e tests.

[GeorgWa]

I think this fits the idea very well apart from how the user triggers it.

Ideally the user should have two Boolean arguments:

transfer_step: true
mbr_step: true 

the first one will prepend a transfer learning to the main search, the second one will append the MBR search to the main search.

The config specification of what should be overwritten looks good.

[GeorgWa]

LGTM

[GeorgWa]

Optional if MBR search, mandatory if TL search

[GeorgWa]

Same here, reuse_quant is not needed.

[GeorgWa]

I use estimator already in the calibration. This is the results of the hyper parameter optimisation. So it could be optimised parameters.

[mschwoer]

good idea, but will be read from stats.tsv anyway