Get an array of results from GPT

[danvk]

#164 follow-up; data update: oldnyc/oldnyc.github.io@2c8e18d

• Change GPT prompt to ask for an array of location candidates.
• Change Coder protocol to return a list of Locatables (could be a generator in the future).

This creates a few new failure modes. A notable one is that the rare photographer who includes their address in the byline becomes a major problem, since I give addresses precedence over place names. I've manually reviewed all addresses that appear in 2+ items or cause a photo to move 1+ km, but there are almost certainly more unique photographer addresses lurking.

• Blacklist many photographer addresses
• Be more careful about stripping directions from street names where they're critical: West Street, South Street, West End Ave, etc.
• Blacklist some "cursed" intersections: Broadway & Amsterdam for now.
• Adds a localturk template for reviewing GeoJSON changes (from diff_geojson.py).

This is a pretty big diff:

Changed: 543
  +geom: 1,321
  -geom: 34

• The drops are a mixed bag. Some are losses, but many were incorrect before, e.g. due to photographer addresses.
• The changes are also a mixed bag.
  • Of the 120 biggest movers (1+ km), 68 (57%) are wins, 30 (25%) are losses and 22 are neutral.
  • On a random sample of 15, it's 1 clear loss (3 small losses), 6 neutral, 3 clear wins (2 small wins)
• The additions seem uniformly good. The first 15 (random sample) that I checked were all wins.

The diff on the truth data doesn't look too hot (+8 wins, +5 losses) but none of the new losses bother me too much. The primary cause is "address pinning", where if I geocode "1234 X St" but the highest address on X St is 789, then Google returns the location of "789 X St." This can happen when the street used to continue farther. I'm hoping that plugging in nyc-streets.geojson from the NYPL will help fix this by giving me more historic intersections.

TODO (this PR):

• Refactor Google block in geocode.py

Remaining follow-up work:

• Blacklist intersections like "59th Street & Central Park"
• Look at what Google actually geocodes. It often makes surprising changes to the street name that we should reject.
• Prioritize highly-specific subjects like "Central Park - The Lake" over intersections.
• West Broadway ≠ Broadway