Fix PUF clone prior weights

[MaxGhenis]

Summary

• give zero-weight PUF clone households meaningful uniform prior mass instead of near-zero epsilon priors
• add legacy SOI Table 1.4A long-term capital gains amount/count targets by AGI bin
• map positive LTCG into legacy SOI target construction and use the LTCG aggregate uprating path for capital gains SOI rows

Tests

• env -u UV_FROZEN uv run ruff check policyengine_us_data/datasets/cps/enhanced_cps.py policyengine_us_data/utils/loss.py policyengine_us_data/utils/soi.py tests/unit/test_enhanced_cps_clone_diagnostics.py tests/unit/datasets/test_enhanced_cps_seeding.py tests/unit/calibration/test_loss_targets.py tests/unit/test_soi_utils.py
• env -u UV_FROZEN uv run pytest tests/unit/test_enhanced_cps_clone_diagnostics.py tests/unit/datasets/test_enhanced_cps_seeding.py tests/unit/calibration/test_loss_targets.py tests/unit/test_soi_utils.py tests/unit/test_refresh_soi_table_targets.py

Pending

• draft until the local ECPS rebuild/calibration diagnostic is rerun with the block-CD distribution storage artifact available

Refs #1139

[MaxGhenis]

Post-merge local calibration impact from the 2024 rebuild diagnostic (ltcg_agi_targets_main, 1,000 epochs):

• Across 3,539 common finite targets, median relative absolute error improved from 9.64% to 5.64%; the share within 10% improved from 51.1% to 68.6%.
• Mean objective fell from 891.70 to 2.46; median objective fell from 0.00843 to 0.00318.
• 1,210 / 3,539 common finite targets worsened (34.2%), even though the aggregate fit improved. The largest worseners were concentrated in high-AGI business/partnership targets and a few demographic/geographic targets: e.g. partnership and S-corp income in AGI $79M+ (299% -> 1,810%), negative_household_market_income_total (609% -> 968%), partnership and S-corp income in AGI $16M-$79M (57% -> 363%), business net profits in AGI $16M-$79M (9% -> 274%), NY AGI count $1M+ (94% -> 203%), and SOI filer count AGI $10M+ (17% -> 91%).
• The PUF clone half now receives material calibrated mass instead of being effectively unused. In the local run the final split was about 74.6M CPS / 91.3M PUF-clone household weight, or 55.0% clone share.
• Capital-gains fit improved enough to make this worth merging, but the top tail is still the main residual issue: summed SOI LTCG AGI-bin amount is about $1.83T vs $1.26T target (1.45x), and the $10M+ LTCG amount is $1.459T vs $450.5B target (3.24x). The $10M+ LTCG count is much closer: 30.4k vs 29.0k (1.05x).

Interpretation: this fixes the prior-weight bug and materially improves the broad calibration surface, but it does not solve the high-end gains amount problem. The next follow-up should focus on the Forbes/PUF top-tail construction and whether those records should be explicit ECPS rows rather than only entering through PUF-derived imputation.