[h-mayorquin]

This is the last peel-off from #4260, the second half of the format change (4 in my review there), and it builds on #4881, so please review that one first. After #4881 the only object arrays left in a zarr SortingAnalyzer are the extension data that is not JSON-serializable, which in spikeinterface means the scikit-learn PCA models of principal_components. I now store their pickle as bytes in a uint8 array, the same way #4881 stores JSON. The content is the same as before and the same as the .pkl files of binary_folder; only the container changes, so a zarr analyzer contains no object arrays at all and #4260 no longer needs a way to write scikit-learn models to zarr. Old analyzers still load.

This is a different approach from #4260, which writes the models with a new helper that saves the estimator's attributes as zarr arrays and JSON. I chose to do what binary_folder already does, which is to pickle them, so both formats store the extension data the same way and we don't add a zarr-only serializer to maintain. I think the models should be stored as plain arrays instead of pickles, but that should change both formats at once, so I will propose it in a separate issue.

I also route the legacy dict and object extension branches through read_encoded_array, so the reader of old object arrays under zarr-python 3 in #4260 has a single place to hook into.

[alejoe91]

@h-mayorquin <3

[alejoe91]

Super!