WIP: port 75y long-run projection engine to microplex-us (rebase on MP base)

[MaxGhenis]

WIP: port the 75y long-run projection engine to microplex-us

policyengine-us-data is being deprecated in favor of microplex-us, and the 75y long-run (CRFB) projection currently builds from the eCPS — it should build from the MP dataset. This PR relocates the engine into microplex_us.longrun ahead of MP promotion so it's ready to rebase the moment MP is promoted. Opening as draft.

Done

• Relocated the 4-file engine (run_household_projection, projection_utils, calibration, ssa_data) into src/microplex_us/longrun/.
• Rewired imports: relative cross-module imports; STORAGE_FOLDER → local longrun/data/ (the only us-data coupling was that one import).
• Ported the SSA Trustees data it depends on (SSPopJul_TR2024.csv, social_security_aux.csv).
• Added the MP-base rebase hook to the dataset registry (mp_2024).
• Library modules (calibration, projection_utils, ssa_data) import + smoke-test clean.

Calibration backends (for reviewers)

The default reweighter is entropy balancing (KL-divergence min vs base weights via scipy.optimize.minimize); the L0 path is microcalibrate / HardConcrete (torch GD + L0). The 75y profile only calibrates age / SS / payroll / TOB — income tax is a free functional, which is why the native model can produce income tax > GDP (see TODO + loop journal iter87).

TODO (post-promotion)

1. Rebase base_dataset on the promoted MP arch-calibrated dataset (set SELECTED_DATASET="mp_2024").
2. Income anchoring — the real fix for the income-tax-over-GDP under-determination: regularize forward weights toward the income-rich MP base + add a CBO long-run income-tax-receipts %GDP target.
3. Refactor run_household_projection.py from a top-level script into an importable def main() entrypoint.
4. Finish lint cleanup of inherited code.

Depends on MP promotion (the 75y base = the promoted MP dataset). Tracks the calibration-fix work in #200 and the under-determination findings in the mp-ecps loop journal.