Save Census geopackage files split by state, then adjust overlaps script

[maria-milosh]

Scripts to modify:

• pipeline/preprocessing/census_geographies.py - currently write_processed() writes a single {year}{level_label}.gpkg. Change it to group by STATEFP and write {year}{level_label}_{statefp}.gpkg per state. The rest of the script are ok as is
• pipeline/preprocessing/overlaps.py currently reads the entire national .gpkg as its first positional arg. It needs to (1) accept a directory (or a pattern) instead of a single file path (2) determine which states the study area geometry intersects (can be derived from the study area boundary) (3) load only the matching {year}{type}{statefp}.gpkg files and take a union if multiple files.
• study_areas.py?
• scripts/run_overlap.sh and scripts/build_study_areas.sh?
• scripts/resolve_config.py

[maria-milosh]

Moving to state level made me re-investigate how block geometries are merged to block population data:

• For 2000, 2010, 2020 miss rate is 0.1%, taking place in FIPS 60 (American Samoa), 66 (Guam), 69 (CNMI Commonwealth of the Northern Mariana Islands), 78 (USVI United States Virgin Islands). We don't download population data for them (but they're included in NHGIS, hence the mismatch). This is okay.

• For 1990: 31.3% national miss rate:

Cause	Count	Share
Zero-population or absent from NHGIS	1,403,445	62.4%
Alpha-suffix split blocks (101A, 204B, …)	453,614	20.2%
BNA tracts (GISJOIN[8:10] == "95")	393,166	17.5%
Space in GISJOIN	2	~0%

The main category is zero-population blocks that exist as geography only. 82% of these have a numbering scheme that is not expected for Census blocks, so they don't match. Block IDs should have 3-4 digit IDs but these have 1. Apparently they come from Enumeration Districts, which is a pre-tract geography used in Alaska, Native American lands, and some other rural areas that hadn't been tracted in 1990. Their blocks get single-digit IDs because enumeration districts used a different internal numbering scheme. They aggregate to 96/97/000xxx tracts.

The NHGIS "blocks" population extract was scoped to census-tracted blocks only, so ED blocks are absent from it
17% look like normal blocks. These must be genuine 0-population blocks, so they're absent from NHGIS population data.

The alpha-suffix blocks are a 1990-specific thing: where a block was split by a political boundary, Census suffixed the parts A/B/C/D in the geometry (to distinguish portions of a collection block that fell into different geographies) but the population table only has the parent block ID. Fixing this would be too cumbersome for very little gain for the moment.

BNA tracts are an analogue of blocks for rural areas and we do not use them (I think most likely the NHGIS extract was scoped to census-tracted areas only?): those areas are outside of our selection of CBSAs (over 100K population) and outside of cities.