Model_based_curation + unitrefine prediction: add unit_ids input

[alejoe91]

The current implementation of unitrefine takes a SortingAnalzyer as input, runs the noise/neural classifier, creates a new SortingAnalyzer with neural units (with remove_units) and then runs the SUA/MUA classification.
This PR allows to run the function (and the model based classificaiton), direclty from computed metrics, preventing the memory copy of the remove_units.

It also simplifies the implementation of some helper function and phy exporter in the model_based_curation

[alejoe91]

Alternatively, extend the model_based_label_units to accept a unit_ids and slice metrics inside

[alejoe91]

Alternatively, extend the model_based_label_units to accept a unit_ids and slice metrics inside

@samuelgarcia modified with your suggestion to optionally pass unit_ids instead aof extending to metrics

[chrishalcrow]

Some of the tutorials on the UnitRefine page use the input_data to pass csvs.

https://github.com/anoushkajain/UnitRefine/blob/main/UnitRefine/tutorial/apply_model/apply_model_using_csv_metrics_files.ipynb

So this might be used in people's code in the wild.

[alejoe91]

good point, reverting!

[alejoe91]

actually, I think metrics is a better name. I'm adding input_data as and a deprecation warning.

[chrishalcrow]

Just want to note that we do this for UnitRefine but not for Bombcell.

[alejoe91]

now I added a set_predictions_as_properties argument to the predict_labels function and it is set to False by the unitrefine functions (so we don't have duplicated columns!)

[alejoe91]

@chrishalcrow I modified the PR after Sam's suggestion. Instead of passing/analyzer or metrics, we can now optionally pass a subset of unit_ids to filter the prediction on. Most of your comments are therefore resolved!