added capability to handle duplicated timestamps in timeseries data.

[pedrofluxa]

This PR fixes the problem of time-series having duplicated timestamps by averaging groups of duplicates (for numerical data) and/or keeping the first duplicated value for non-numerical data.

A very simple unit test was added as well.

[paxcema]

I think the failing test is due to the input DF no longer being equal in size to the output DF. This is fine, but the test needs to change.

[paxcema]

While overall the approach is nice, there is one important omission: the input df for clean_timeseries may contain several different groups, each of them with duplicate (but valid!) measurements.

We need to handle this differently for the case when tss['group_by'] is a list. I suggest moving the new procedure into something like deduplicate_single_series, then use this for each existing group.

As it stands, this breaks grouped series flows (e.g. lightwood's grouped time series unit tests), so we shouldn't merge yet.

[paxcema]

Added a test that replicates the original Lightwood test which caught a bug in the original implementation of this PR 👍