Refactor base indicator classes

[aulemahal]

Pull Request Checklist:

• This PR addresses an already opened issue (for bug fixes / features)
  • This PR fixes Frequency analysis misleadingly characterized as resampling indicator #1263
• Tests for the changes have been added (for bug fixes / features)
  • (If applicable) Documentation has been added / updated (for bug fixes / features)
• CHANGES.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?

Move around some of the base indicator classes code. Split the missing values handling from the resampling handling. This way, indicators like return_level can reduce the full time axis and still perform some missing values handling.

Does this PR introduce a breaking change?

No, not yet.

Other information:

@huard @RondeauG, if I am not mistaken, return_level and fit currently have missing values handling disabled because it was impossible with the previous classes to have both this and full reduction of "time". Now that it is possible, should we activate it ?

Similarly, stats had the "Any" missing method forced. Is there a reason for that ? Here, I removed the argument, meaning it will use "from_context".

[RondeauG]

if I am not mistaken, return_level and fit currently have missing values handling disabled because it was impossible with the previous classes to have both this and full reduction of "time". Now that it is possible, should we activate it ?

Similarly, stats had the "Any" missing method forced. Is there a reason for that ? Here, I removed the argument, meaning it will use "from_context".

Absolutely ! For all 3, having full control on the missing argument would be a great quality-of-life change.

[coveralls]

coverage: 90.674%. first build when pulling 71c52a7 on indcls-fullreduce into 36dcbdf on master.

[huard]

Issue CSHS-CWRA/RavenPy#307 is probably due to the fact that this PR made fit a subclass of ReducingIndicator. It was previously a subclass of Indicator, which has no missing value logic in it.

So the generic fit indicator essentially went from skipping missing value checks, to apply the missing algorithm from context (typically any). I think we should have highlighted this more clearly in the release notes.