Add get_equidistant_timeseries() method

[dbrakenhoff]

Short Description

Add method for getting equidistant timeseries. Could be useful sometimes, but especially necessary for the stoffer_toloi test (see #336). This test was using pd.Series.asfreq to get an equidistant timeseries of the noise but in certain situations this dropped a lot of data, resulting in a calculation being performed on a subset of the noise. This new method addresses that, but it does perform some nearest interpolation. The old method did not interpolate, and only sampled existing data points that occur at the user-specified frequency.

I'm still looking for some input, as I'm not entirely sure whether we would prefer the rather complicated method I added, or something simpler using pandas methods with a few limitations as shown in the notebook. Also not sure if we want a notebook showcasing this method necessarily. So suggestions/feedback welcome on this PR!

Checklist before PR can be merged:

• closes issue [ENHANCEMENT] improve reindexing to equidistant timeseries in stoffer_toloi #336
• is documented
• PEP8 compliant code
• tests added / passed
• Example Notebook (for new features)
• API changes documented in Release Notes

[raoulcollenteur]

Stoffer-toloi is for missing data, whereas many time series are irregular. One way to deal with the irregular time steps is to interpolate and get missing data time series. Another approach would be to conclude that the stoffer-toloi is not applicable. Yet another method would be is to select a smaller time step through the freq argument (perhaps this is the best).

I would prefer to make the interpolation in the stoffer-toloi test optional through a keyword-argument in that method, and give users control over that. The current PR makes a fundamental change without users being aware which should IMO be prevented.

[dbrakenhoff]

I'm okay with making the behavior optional, but if we use the asfreq() method as an alternative, I think we need to add a warning when the resulting equidistant noise series is missing more than some percentage of the number of original noise values?

And to be clear, the current implementation in pastas will silently drop a bunch of noise values if the original timeseries timesteps are not equidistant, especially if the timestamp of the first observation is different from others in the timeseries. The method get_equidistant_series() attempts to prevent this loss of data by creating an equidistant series that takes as much data as possible from the original noise series. It does shift observations in time (slightly) to create this equidistant timeseries. I think the latter is preferable in situations when asfreq() would drop a lot of values. opti

Update: Also I propose changing the asfreq() method to check which time offset is the most common:

t_offset = ps.utils._get_time_offset(series.index, freq).value_counts().idxmax()
new_idx = pd.date_range(
    series.index[0].floor(freq) + t_offset,
    series.index[-1].floor(freq) + t_offset, 
    freq=freq
    )
series.reindex(new_idx)

This avoids losing a lot of data if the first observation happens to have a different time offset than the other observations.

EDIT: improve snippet using floor() instead of normalize()