Use person-weighted groups for income-decile impacts

[anth-volk]

Problem

UK Chat delegates decile statistics to policyengine.py, but the upstream
fallback grouping used pandas.qcut. That ignored MicroSeries weights and
assigned deciles by raw survey rows.

For household income distributions, computed groups should be person-weighted:

MicroSeries(
    income,
    weights=household_weight * household_count_people,
).decile_rank()

Resolution

Keep the weighting implementation in policyengine.py, not UK Chat:

• Default decile impacts to household net income and person-weighted groups policyengine.py#486 provides shared person-weighted fallback
  groups for DecileImpact and IntraDecileImpact.
• UK Chat pins the exact upstream commit until the first containing release is
  available.
• UK Chat continues to pass its three public concepts explicitly:
  household_net_income, equivalised_hbai_net_income, and wealth.
• UK Chat does not materialize decile assignments or manipulate survey weights.
• Wealth impacts continue to pass the precomputed
  household_wealth_decile.

No policyengine-uk change is required.

Acceptance criteria

• UK Chat contains no local weighted-grouping implementation.
• Ordinary and explicit equivalised-HBAI income concepts use upstream computed
  person-weighted groups.
• Wealth grouping remains explicit and precomputed.
• The upstream dependency is pinned immutably and covered by backend tests.
• Prompts, tool descriptions, and runtime guidance reflect the calculation
  boundary.

[vahid-ahmadi]

Review: this is well scoped and keeps the calculation boundary in the right repository. PR #217 implements the stated upstream delegation and explicit three-concept contract; closure should wait for that draft to be rebased, the unrelated preview-deploy changes to be split or separately justified, and the pin to move from 5.0.0 to the latest reviewed 5.0.1.