Scale preprocessing

**Input**: npz file
**Output**: pre-processed npz file

Historically, preprocessing has been very light: some intensity normalization and converting to float64 type.

The problem is, for a 1TB image, we can’t just load it into memory. We estimate max GPU to be a few hundred GBs.

Still, we think the parallelization is “straightforward” (e.g. find min/max and/or histogram per tile then normalize each tile).