Replace CD stacking with block-based county FIPS filtering for NYC.h5

[baogorek]

Problem

NYC.h5 is currently built using "congressional district stacking" — filtering to 13 CDs that overlap NYC, then probabilistically scaling weights by P(NYC county | CD). This approach is deprecated because:

1. CDs are redrawn every decade — the hardcoded NYC_CDS list of 13 CDs is fragile
2. NYC is not a collection of CDs — CDs straddle NYC boundaries, requiring probabilistic weight scaling
3. We now have census blocks — GeographyAssignment.county_fips gives us 5-digit county FIPS derived from block_geoid[:5]

Solution

Replace CD stacking with a direct county FIPS filter:

NYC_COUNTY_FIPS = {"36005", "36047", "36061", "36081", "36085"}

Each clone IS or IS NOT in NYC based on its assigned block's county — no probabilistic scaling needed. This is simpler, more correct, and doesn't depend on congressional district boundaries.

Changes

• Add county_fips_filter parameter to build_h5() that zeros out weights for clones outside target counties
• Update build_cities() to use county_fips_filter=NYC_COUNTY_FIPS instead of cd_subset + county_filter
• Remove NYC_COUNTIES (enum name set) and NYC_CDS (13 hardcoded CD codes)
• Remove now-unused get_county_filter_probability() and get_filtered_block_distribution() from block_assignment.py
• Update modal_app/worker_script.py accordingly