Additional tweaks to improve performance of household inference

[dehall]

This PR introduces yet more tweaks to improve performance of household inference in households.py.

Summary:

• Only read the pii columns required
• Keep the "exploded address" columns in a separate DataFrame so we can delete the whole thing once those columns are no longer needed
• Dump the household matched pairs to a file so that we can restart from there if the process runs out of memory
• Add a --pairsfile arg to specify that pairs file to restart from there
• Write the household_pii and mapping files at the same time, don't just store the household_pii file in an array to write later
• Add some additional debug statements
• Keep using MultiIndexes wherever possible rather than converting to lists of tuples because the performance seems to be better all around
• Add the [extras] to the textdistance dependency because it includes additional libraries that speed up, ex, jarowinkler. In my testing this was about a 25% speedup.
• Delete objects and aggressively GC when they are no longer needed