Use stable sort when sorting values in forecaster

[natl]

When Prophet prepares a dataframe for fitting or predicting, it runs a sort on the ds column.

This changes the sort algorithm from the pandas default sort to mergesort, which is a stable sort.

This is important in situations where multiple regressors are present and so the fitted dataframe is not unique by date.

Users may do something like the following:

df = <some dataframe>
# Example df
# ds            x   y   
# 2020-01-01    0   3.2
# 2020-01-01    1   4.1
# 2020-01-02    0   3.7
# 2020-01-02    1   4.3

m = Prophet()
m.add_regressor('x')
m.fit(df)

predict_df = m.predict(df)
df['prediction'] = predict_df['yhat']

In m.predict(df), there is a call to df.sort_values('ds'). As this is not a stable sort, sometimes the x column above can get switched around, even if df is sorted. A stable sort addresses this.

This partially addresses #2322 if you construct the input correctly.

[tcuongd]

Thanks! I'm not sure Prophet was intended to be used the way you described (i.e. as a regression engine). Typical guidance is to have one regressor value per date. But I guess your example does just work out of the box, so this change makes sense to help people using it in this way. Only other consideration I'd have is performance but given time series datasets are usually too small for this to matter + this is a training time operation, I don't think it's a big deal.