Use shared WEEKS_IN_YEAR and local RNG in consumption imputation

[MaxGhenis]

Summary

datasets/imputations/consumption.py had two sub-bugs (finding U6):

1. Weekly-to-annual conversion used a bare * 52 while the sibling FRS loader uses WEEKS_IN_YEAR = 365.25 / 7 ≈ 52.1786. That ~0.34% gap systematically underestimated annualised LCFS consumption relative to FRS income and skewed VAT/energy imputation targets.
2. Two np.random.seed(42) calls mutated the process-wide RNG state, so any unrelated numpy random consumer that ran after consumption imputation silently changed behaviour.

This PR:

• Promotes WEEKS_IN_YEAR to module scope in datasets/frs.py (with a docstring) so siblings can import one canonical value, and swaps consumption.py to use it for both the household and person annualisation passes.
• Replaces both np.random.seed(42) calls with rng = np.random.default_rng(42) and routes the draws through the local RNG, keeping the seed at 42 for bit-for-bit fingerprint compatibility.
• Adds test_consumption_weeks_rng.py asserting the imported constant matches, the module source no longer contains np.random.seed(, and re-importing consumption does not perturb the global RNG state.

Test plan

• uv run pytest policyengine_uk_data/tests/test_consumption_weeks_rng.py -q passes.

Finding U6 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.)