[codex] Add LIHEAP targets to local calibration DB

[MaxGhenis]

What changed

This PR adds LIHEAP recipient-household targets to the shared national calibration target pipeline and makes the legacy national EnhancedCPS reweighting path consume those targets from policy_data.db.

It does three things:

• adds FY2023 and FY2024 national LIHEAP recipient-household targets to policyengine_us_data/db/etl_national_targets.py
• extends the national target ETL so conditional targets can load non-person_count variables, which LIHEAP uses as household_count
• adds a legacy build_loss_matrix() bridge in policyengine_us_data/utils/loss.py that reads active LIHEAP targets from storage/calibration/policy_data.db

It also clarifies in the repo docs that there are still two calibration tracks in-tree:

• the current unified policy_data.db calibration pipeline
• the older national-only EnhancedCPS / build_loss_matrix() path kept for legacy reproduction

Why

LIHEAP was already present in CPS/SPM reported values, but it was not represented as a calibration target in either:

• the shared national target database pipeline, or
• the legacy local EnhancedCPS reweighting path

That meant we could not calibrate LIHEAP recipient counts in either path, despite having CPS/SPM baseline LIHEAP receipt in the data.

Scope note

This PR calibrates LIHEAP recipient households only. It does not calibrate LIHEAP dollars.

That is intentional for now: the CPS/SPM concept is a pooled annual reported energy-assistance amount (spm_unit_energy_subsidy_reported), while LIHEAP administrative dollars are split across heating, cooling, crisis, and weatherization. Recipient-household counts are a cleaner crosswalk than national dollars.

Follow-up issue: #689 tracks options for national LIHEAP dollar targeting.

Impact

• make database / etl_national_targets.py now loads national LIHEAP household-count targets into policy_data.db
• local legacy build_loss_matrix() runs can pick those LIHEAP targets up from the DB when present
• the docs now state more explicitly that target changes sometimes need to touch both calibration tracks

Validation

• python3 -m py_compile policyengine_us_data/db/etl_national_targets.py policyengine_us_data/utils/loss.py

I also attempted:

• python3 -m pytest tests/unit/test_etl_national_targets.py -q

That pytest invocation did not run in this shell because the environment is missing torch, which is imported from the repo test setup.