Move the Base.save to Sorting.save() and Recording.save()

[samuelgarcia]

Move saving logic to extractor classes.

For historical reason, lazyness and original bad design, the BaseExtractor.save() was a terrible mess for:

• save/load properties
• annotation
• dump provevance
• create the "si_folder.json"
• ...
  It was using some kind of "hook" to sub classes and the logic was almost impossible to follow.

I tried to move the logic save()/load() to classe (BinaryFolderRecording, ZarrRecordingExtractor, SharedMemoryRecording, ZarrSortingExtractor, NpzFolderSorting, NumpyFolderSorting).
The code looks quite strange at soe places but at least it is easier follow.

@alejoe91 : the time vector handling is not done correctly.
But your PR #4689 should be rebase on this one I think. Lets do it togoether.

[alejoe91]

@samuelgarcia fixed some bugs in preprocessing, but there are more. Do you want me to fix it?

[alejoe91]

@samuelgarcia ported changes from #4689 (timestamps handling + removing metadata_folder) and cleaned up code and docs.

One main question: in #4689 , I added an extra _in_reduce mechanism to the to_dict/_extra_metadata_to_dict stack. This was only set to True by the __reduce__, and propagated times kwargs only in this case.

Here I propose a modified version: if the user "messes" with timestamps, this is propagated to dict no matter what. Note that:

• this is needed when saving in parallel (timestamps need to be saved!)
• the final si_folder.json will NOT have this information, since it sets include_extra_metadate=False (added also include_annotations/include_properties to False)

Let me know what you think!

[samuelgarcia]

I think we should limit the complexity of BinaryRecordingExtractr to load only one or sevreal binary files.
And keep the complexity only in BinaryFolderRecording other we need to maintained a 2 classes design.