Preserve partnership loss support in donor imputation

[MaxGhenis]

Summary

• Add a signed zero-inflated support family for variables with real negative/zero/positive donor support.
• Classify person- and tax-unit partnership/S-corp income as signed zero-inflated so donor imputation remains support-aware without forcing negative values to zero.
• Keep true positive-only variables in nonnegative_vars, and keep signed zero-inflated variables in the ZI-QRF zero gate.

Why

Raw PUF has partnership/S-corp loss support, but the Gate-2 candidate H5 had zero negative partnership_s_corp_income rows. eCPS has negative rows and nonzero partnership/S-corp loss targets.

The cause is that Microplex metadata marked partnership/S-corp income as ZERO_INFLATED_POSITIVE, which put it in nonnegative_vars; both donor imputer backends then clamp negative predictions with np.maximum(..., 0).

Validation

• uv run --python 3.13 --extra dev pytest tests/test_variables.py tests/pipelines/test_us.py -k 'partnership_income_semantics or signed_zero_inflated_donor_vars or zi_qrf_backend'
• uv run --python 3.13 --extra dev ruff check src/microplex_us/variables.py src/microplex_us/pipelines/us.py tests/test_variables.py tests/pipelines/test_us.py
• uv run --python 3.13 --extra dev ruff format --check src/microplex_us/variables.py src/microplex_us/pipelines/us.py tests/test_variables.py tests/pipelines/test_us.py

Notes

This does not touch the active rent/childcare PR lane. It removes a support-family mismatch; it does not add any target bucket cap or record-level adjustment.