Finalize the MTL logic for reprocessed tiles

[geordie666]

This PR extends #774 by including the actual, final logic for reprocessing tiles. This logic is to:

• Assemble all previous observations that touch the set of reprocessed tiles including any "new" observations from the reprocessed tiles.
• Find the most recent observations for unique combinations of TILEID + TARGETID and store these in a redshift catalog (which I'll refer to as the "all-zcat" below).
• Determine the order in which tiles were originally processed and loop through the tiles in this original order:
  • Start with the UNOBS state for a given TARGETID.
  • Update the UNOBS state (progressively, in tile order) using entries from the "all-zcat" for a given TARGETID.
  • Hence, recover the final state for each TARGETID with the reprocessed observations substituted for the "orginal" observations.
• The full progression of states for each TARGETID is then appended to the ledgers, including BAD observations (which, previously, never appeared in the ledgers).

This PR also includes:

• During reprocessing, TIMESTAMPs are mocked up in advance for speed.
  • i.e. instead of having to delay the code by 1 second for each tile so as to ensure a unique TIMESTAMP.
• The numobsfromzcat flag is deprecated as a user-specified option when executing run_mtl_loop.
  • as we have always, in actuality, retrieved NUMOBS from the ledgers themselves rather than the redshift catalog.
  • it seemed increasingly fragile to retain an option we've never used.

[geordie666]

A full set of example reprocessed ledgers is available in /global/cscratch1/sd/adamyers/fakesurveyops4.