Explore per-algorithm data transformation

See `1.compress_given_z.py`. Right now, gene expression data is loaded and preprocessed once, then multiple algorithms can be run on the preprocessed data.

With new algorithms being added this isn't always ideal: for example, NMF only works on data with positive values (so z-score preprocessing won't work), but PLIER only works on data with no zero values (so zero-one rescaling won't work)

It would be good to find a performant way to use different preprocessing approaches based on which compression algorithm is being run.

Start with 1.compress_given_z.py and trace where gene expression data is loaded, preprocessed, and passed to each compression algorithm. Compare the preprocessing constraints described for NMF and PLIER, then define a performant per-algorithm transformation flow; done means each algorithm receives compatible data without unnecessary preprocessing work.

Written by the indexing model from the issue text.