National generalization of the ACS SLD packages (all states, ~6.8k districts)

[MaxGhenis]

Follow-up to the Utah-pilot SLD packages (branch census-acs-sld-2024, populace#625). Verified mechanics for going national, so the generalization is mechanical once the pilot shape passes judge review:

• Fetch: the API rejects in=state:* for SLD hierarchies but accepts comma-listed FIPS (in=state:01,02,...) — one canonical national witness per (table, chamber) works, same single-file shape as the CD packages. 50 states + DC; PR excluded per the populace state spine.
• Scale (measured from the pilot's compact record-set shape — rows = districts with geography overrides, measures = columns with constraints): ~1,950 SLDU + ~4,833 SLDL districts → roughly 2.3MB YAML for the biggest package (S0101 lower) and ~3-4MB for the biggest witness; totals ~12MB YAML + ~12MB witnessed JSON across the six packages. CD-package-proportionate, no harness change needed.
• Suppression handling to add: nationally, B19013 medians (and possibly sparse B19001 cells) will carry ACS suppression markers for thin districts. The pilot generator hard-fails on suppression; the national generator needs the skip-with-record path — suppressed cells excluded from the package with an explicit exclusions list in the package (never silently absent).
• Special cases already structural in the pilot: NE unicameral and DC wards (no lower chamber), multi-member districts as Census tabulates them; NH floterial representation should be verified against the witnessed rows when the national fetch lands.
• Ordering note: witness row order is the API's (sorted by state then district in practice); the generator asserts sortedness per state and derives row numbers from the parsed stream, so ordering drift is a loud failure.

Blocked behind: the pilot PR's judge review (shape adjudication) and the pending publish-raw R2 run for the pilot manifests.