Migrate to selection-first approach in classify_altloc_selections.py

[k-chrispens]

Currently classify_altloc_selections.py detects altloc selections and builds all pairs of altlocs before applying selections and the classifications, rather than using the input selections and then building an AtomArrayStack with the altloc selections using grid_search_utils.map_altlocs_to_stack, which would be cleaner and probably faster: see #223 (comment)