Fix SPI __main__ crash, parameterise marriage allowance, seed age RNG

[MaxGhenis]

Summary

policyengine_uk_data/datasets/spi.py had four small but load-bearing bugs (bug-hunt finding U2):

• The __main__ block called create_spi(path, year) with two positional args, but the signature required three — running the module directly crashed with TypeError: create_spi() missing 1 required positional argument.
• Marriage allowance was hardcoded to 1_250 (2020-21), so builds for later fiscal years produced the wrong cap for every row where MAIND == 1.
• np.random.rand was used unseeded, so age imputation drifted non-deterministically across builds.
• Unknown GORCODE values silently fell through to SOUTH_EAST, quietly distorting regional income totals.

This PR:

1. Makes output_file_path optional so two-arg calls and the existing __main__ entry point both work.
2. Reads the Marriage Allowance cap from policyengine-uk parameters (personal_allowance × marriage_allowance.max, rounded up by rounding_increment — the same formula used in policyengine-uk's marriage_allowance variable) keyed on the fiscal year.
3. Seeds age imputation via np.random.default_rng(seed) with seed=0 by default.
4. Defaults unknown region fallback to "UNKNOWN"; legacy "SOUTH_EAST" behaviour is still accessible via unknown_region=....

Adds policyengine_uk_data/tests/test_spi_build.py covering each of the four fixes with a lightweight fake SPI .tab file.

Test plan

• uv run pytest policyengine_uk_data/tests/test_spi_build.py passes.
• Signature test catches any regression that removes the __main__-friendly default on output_file_path.
• Deterministic-RNG test catches any regression that switches back to the global np.random state.
• MA cap test asserts 2020-21 stays at £1,250 and later years diverge (HMRC-published £1,260 for 2025-26).

Finding U2 from the bug hunt.

[MaxGhenis]

Self-review: APPROVE. See /tmp/bug-hunt/reviews/uk-data-reviews.md for detail. (GitHub blocks self-approve, so leaving as comment.)