Optimization

[raymondwjang]

A few standouts:

1. Residual calculation: especially with hundreds of components, calculating slices for all of them takes a while. I think we can just set the coordinates to zero instead, assuming NMF performed correctly.
2. component correlation: the merge step probably takes a while, (have not profiled yet.) with calculating overlaps of all footprints and then correlation between all the traces. let's move to centroid method, and calculating trace correlation only against close components