NA handling in subdaily models etc

[davidorme]

Describe the bug

The main PModel code is not bothered about np.nan values in inputs - there will be a gap in predictions but it all works. However, the memory_effect and fill_daily_to_subdaily methods in FastSlowPModel iterate over time series and get derailed by missing values.

There are various ways this could be handled - and these are most easily dealt with by removing np.nan values before fitting models. Then users can decide what to do rather than having baked in approaches. However, the methods should then look for np.nan in inputs (perhaps excluding leading and trailing np.nan) and raise a ValueError.

Expected behavior

• FastSlowPModel fails with missing data and reports the need to fill gaps in inputs.
• Docs for the approach make this clear and point to useful solutions like xarray.ffill or scipy.interpolate etc.

[davidorme]

This is partly resolved by #78 closing #77, but this doesn't document the expectation that forcing variables should not have missing data.

[MarionBWeinzierl]

Could use None if using xarrays, but it's slow.

Helper function to check about NAs in the functions that care about NAs, but not in others?

Use a specific error check?

[mondus]

Good description of this here: https://www.residentmar.io/2016/06/12/null-and-missing-data-python.html

[davidorme]

This is now covered by #78 and also the more extensive documentation and handling of missing data arising from allowing partial start and end days with the subdaily model in #194.