Add microsimulation smoke test against unpinned latest data

[MaxGhenis]

Summary

Catches silent model/data skew at the point the enhanced FRS dataset is republished on HuggingFace, not after a release. Exercises whatever is currently on HF main (unlike conftest.py which pins to an older version) and asserts plausibility bounds on:

• UK weighted population and household/benunit counts
• is_parent weighted population (>10M) — catches the defaulting-to-zero failure introduced by removing the inferred formula in Remove inferred is_parent formula #1595
• Universal credit aggregate in £55–£95bn range around the OBR target
• State pension / child benefit / pension credit floors
• extended_childcare_entitlement_eligible reaching >500k benefit units

Why

Between 2026-04-15 14:04 UTC (when #1595 removed the inferred is_parent formula) and 2026-04-16 14:17 UTC (when uk-data 1.50.3 published an FRS build with an is_parent column), any run of policyengine-uk main against the prior dataset silently produced:

• is_parent = 0 everywhere
• extended_childcare_entitlement_eligible = 0 everywhere
• Universal credit aggregate ≈ £51.7bn (vs OBR target ~£74bn), because capital-limit logic combined with stale savings imputations over-triggered disqualifications

None of that surfaced in CI because the existing microsimulation-marked tests pin to dataset version 1.40.3 via conftest.py. This adds a parallel, unpinned test that uses whatever is on HF main — so the next time model code and data pipelines drift apart, CI notices.

Verification

Local run against fresh 1.50.3: all 5 tests pass.

Local run against the stale 1.45.8 snapshot from before the fix:

PASS: test_population_totals_are_plausible
FAIL: test_is_parent_is_populated — is_parent weighted total 0 is too low
FAIL: test_universal_credit_aggregate_in_range — UC £51.7bn outside £55–£95bn
PASS: test_core_benefits_are_nonzero
FAIL: test_childcare_entitlement_populated — extended_childcare_entitlement_eligible total 0

i.e. the three failure modes the April 15–16 window actually exhibited all trip.

Scope

• policyengine_uk/tests/test_latest_data_smoke.py — 5 new tests, all marked microsimulation so they skip unless HUGGING_FACE_TOKEN is present.
• changelog.d/latest-data-smoke.added.md — changelog fragment.
• No production code touched.

Test plan

• uvx ruff format --check / uvx ruff check clean
• All 5 tests pass locally against fresh HF main (1.50.3)
• 3 of 5 tests fail as expected against stale 1.45.8 snapshot
• CI Test job passes