Improve Per-Band Support in Ensemble.batch

[wilsonbb]

It's pretty typical for a user to want to perform per-band analysis via Ensemble.batch. Currently, a user can implement themselves in their custom function via array masking, but this is a) poor user experience b) gives the user opportunity to make an error.

Currently, Ensemble.batch's on parameter could be used to allow the user to specify that the results should be grouped by both object ID and band. This yields an output table with a MultiIndex which (in my attempts at playing around with TAPE) initially works but eventually produces an error in the computation graph (unsurprising since Dask lacks MultiIndex support).

Possible solutions include:

• @dougbrn has suggested that we could consider using a pivot table in Ensemble.batch similar to what we do in Ensemble.calc_nobs
• Provide a specific per_band boolean parameter to Ensemble.batch that will take a user's output columns and create one per-each observed band (though we probably want to do some variation of the above anyway).

[hombit]

I like this idea, it would simplify feature extraction support. However, I see some discussion points here. Would we like user to select a list of bands? Do we really want it in batch?

1. I don't really understand the first proposal...
2. It would make batch to be quiet complicated, especially in meta specification by user and return types. I would think about alternatives, for example having a separate batch_per_band or something like that.