Handle nodata or other ignore values in raster data sets

[hannah-rae]

Rasters often have nodata values that should be ignored in analysis (e.g., edges missing data due to registration). These pixels should be given a specific sentinel value (e.g., -9999) instead of being scored.

[hannah-rae]

@bdubayah Have you made this change for the raster pixel loader too? Are you planning to PR/merge or are you still making changes?

[bdubayah]

Still need to do it for raster pixels. I think we also need to decide the merge strategy for overlapping patches - whether to stick with mean or do max instead. I'm going to try with max for the vtfs experiment and see how it works out.