Consolidate economic impact analysis into shared compute functions

[anth-volk]

Problem

The economic impact analysis logic (decile impacts, budget summary, program statistics, poverty, inequality) is currently duplicated between the US and UK analysis.py files, and depends on country-specific string checks scattered throughout the code. This makes it hard to maintain, test, and extend to new countries.

Solution

Introduce shared, country-agnostic compute functions and a CountryConfig frozen dataclass that encapsulates all country-specific parameters:

• CountryConfig (outputs/country_config.py): Frozen dataclass with US_CONFIG and UK_CONFIG singletons holding income variables, program definitions, budget variables, poverty variables, and inequality settings.
• BudgetSummaryItem + compute_budget_summary() (outputs/budget_summary.py): Computes baseline/reform/change totals for budget variables using Aggregate.
• compute_program_statistics() (outputs/program_statistics.py): Computes per-program statistics (totals, counts, winners, losers) using ProgramStatistics/ProgrammeStatistics.
• compute_decile_impacts() (outputs/decile_impact.py): Computes decile-by-decile impacts from already-run simulations.
• PolicyReformAnalysis (outputs/policy_reform_analysis.py): Unified result container assembling all analysis outputs.

Each country's analysis.py calls these shared functions with its config, eliminating duplication.

Additional fixes included

• Fix DecileImpact.run() bare next() → next(..., None) + ValueError (prevents StopIteration crash)
• Fix Aggregate and ChangeAggregate with same pattern (already done in prior commits)
• Move inline imports to module top-level in us/analysis.py
• Improve exception logging in compute_program_statistics with exc_info=True
• Add model_rebuild() calls for ProgramStatistics, BudgetSummaryItem, and PolicyReformAnalysis
• Add 11 new tests covering CountryConfig, compute_decile_impacts, DecileImpact error handling, and ChangeAggregate error handling

[MaxGhenis]

Re-scoping after the PR #260 review: most of the consolidation this issue asked for has happened — both countries' economic_impact_analysis are now thin orchestrators over shared convenience functions in policyengine.outputs (decile impacts, intra-decile, poverty, inequality, LSR, cliffs; see docs/impact-analysis.md "Composing manually").

The residual duplication is narrow and concrete:

• _validate_program_statistics_config — us/analysis.py:104 and uk/analysis.py:112, identical except the country word and the US_PROGRAMS/UK_PROGRAMS reference
• _format_missing_program_variables and the config-error message builders — identical
• the program-statistics build loop inside each economic_impact_analysis — identical but for the programs dict

Plan: extract validate_program_statistics_config(programs, baseline, reform, country_label) and build_program_statistics(programs, baseline, reform) into policyengine.outputs.program_statistics, call from both country files. Explicitly NOT adopting a strategy protocol or a unified cross-country result container — the country containers deliberately diverge (UK wealth-decile outputs), which is what overtook PR #260's design.

Fresh PR to follow shortly.