Better resample

[aulemahal]

Pull Request Checklist:

• This PR addresses an already opened issue (for bug fixes / features)
  • This PR fixes Resample to seasons - exclude seasons with missing months #80 and fixes Period-length ponderation in means #93.
• (If applicable) Documentation has been added / updated (for bug fixes / features).
• (If applicable) Tests have been added.
• This PR does not seem to break the templates.
• HISTORY.rst has been updated (with summary of main changes).
  • Link to issue (:issue:number) and pull request (:pull:number) has been added.

What kind of change does this PR introduce?

• Extends resample to explicitly cover the case where the initial_frequency is coarser than "W", meaning that it has a non-uniform length.

  • Weights are the proportion of days of the total period contributed by each timesteps. Ex: MS to YS => weight = daysinmonth / 365 .
  • "weighted resample" is activated for methods : "wind_direction", "mean", "std", "var" and "median". Those are the methods from DataArray.resample for which a weighted version makes sense.
  • Implemented a generic version that uses resample(...).map(....weighted().op()) , with a special case for "median" since that's not a weighted op. Mapping a function is less performant, especially since flox can't be used.
  • Implemented a shortcut for method='mean' that simplifies to a simple resample().sum(), meaning it should be much more performant and able to use flox. Without dask and on the air dataset, I had a 3x speedup.

• Extends resample to handle missing values with new missing arg:

  • mask : Mask incomplete periods (as in missing timesteps)
  • drop: Drop incomplete periods (same)
  • dict with method=<xclim missing method> : Mask periods failing specified xclim missing check.

Does this PR introduce a breaking change?

No.

Other information:

[RondeauG]

LGTM!

[mccrayc]

Looks good, thanks!