Feat timeseries

[jkeupp]

Motivation

bofire is lacking support for properly handling experimental data that is recorded as a "timeseries"; e.g. for the same experiment, at e.g. 10 different timepoints the output values are determined. Whenever this is the case, cross-validation needs to properly handle this as to not split this data so part of a single trajectory goes into train and test at the same time. This PR adds a flag
is_timeseries to numerical input feature, which if enabled for a feature requires an _trajeftory_id in experiments to define which experiment it belongs to.
cross_validate then automatically performs the appropriate GroupKFold split instead of simply a random split.
There is also a conveniece util function to automatically create the _trajectory_id column from a dataframe by checking for equivalent values in all other input features (up to some eps).

Have you read the Contributing Guidelines on pull requests?

Yes

Test Plan

• The respective cases when cross_validate should behave differently are tested explicitly.
• specs tests were fixed to accomodate the new input feature is_timeseries flag.
• timeseries utils tests were added.

[jduerholt]

Just tell me when it is ready for review ;)

[jkeupp]

hmm, yes... We left off by disucssion whether the group split column would be also make sense stand-alone without the time-series feature in itself. I think that makes sense, but i've not yet updated any code. it'll probably take a while.