Fix ACA PTC state calibration targets

[daphnehanse11]

Summary

• Correct the IRS SOI geography mapping for aca_ptc from 85530 to 85770 so it targets total premium tax credit instead of Additional Medicare Tax.
• Add aca_ptc_state.csv from IRS SOI Historical Table 2 (N85770/A85770) plus a refresh script and loader that applies the checked-in 2022-to-target-year ACA PTC multipliers.
• Use the SOI/uprated ACA PTC targets in the enhanced CPS takeup override, legacy loss matrix, and ACA diagnostic CLI, while keeping CMS enrollment targets for APTC person counts.
• Store local-H5 county enums as integer indices in publish_local_area.py, matching entity_clone.py, so non-ASCII counties like DOÑA_ANA_COUNTY_NM do not fail dataset loading.

Fixes #805.

Diagnostics

Baseline all-state diagnostic on the staged 1.89.1 state files showed Florida as the only major enrollment impossibility: target 4.47M APTC people vs 2.96M modeled positive-PTC potential people (-33.8%). New Mexico failed to load because the staged H5 stored DOÑA_ANA_COUNTY_NM as UTF-8 bytes.

After applying the SOI/uprated target path, the 17-state problem slice shows NY/DC/MN are no longer the scary cases under assigned takeup spending:

state	enrollment error	assigned spending error
NY	+0.1%	+0.3%
DC	-0.3%	+2.4%
MN	-0.0%	+11.0%
FL	-33.8%	-55.4%
TX	-0.0%	-34.6%
GA	+0.0%	-30.7%

So the remaining work is concentrated in real modeled composition/potential gaps, especially FL/TX/GA, rather than NY/MN/BHP target-definition weirdness.

Tests

• .venv/bin/ruff check ...
• .venv/bin/python -m pytest tests/unit/test_aca_ptc_targets.py tests/unit/calibration/test_publish_local_area.py tests/unit/calibration/test_entity_clone.py tests/unit/test_refresh_aca_ptc_state_targets.py tests/unit/test_enhanced_cps.py tests/unit/calibration/test_loss_targets.py tests/unit/test_etl_irs_soi_overlay.py -q
• uv run pytest tests/integration/test_build_h5.py -q

[MaxGhenis]

Fixed the Towncrier fragment path and rechecked the rerun. All PR checks are green, including integration.