Reduce repeated setup in expensive Rest tests

[hua7450]

Related to #9904.

What and why

The Rest tests repeat expensive setup for identical marginal-tax-rate scenarios and rebuild a baseline model already available to the federal-credit tests. Reuse copied, read-only results for each marginal-rate cap, and supply the existing module baseline to each independently constructed credit reform.

• Marginal-rate coverage: 10 simulations become 7. Only output arrays are cached; each distinct cap and roommate scenario still has its own simulation and reform.
• Federal-credit ordering: 3 explicit full baseline builds become 1. Each reform retains its own parameter tree and variable registry.
• Preserve all 29 cases, their parameterizations, populations, years, assertions, and existing Rest process groups.

Local cost measurements

Test file / Rest group	Full-process time, before → after	Peak RSS, before → after
test_marginal_tax_rate_coverage.py / heavy-b	118.5 → 76.3 seconds	2.62 → 2.27 GiB
test_federal_credit_limit_order.py / remaining	134.3 → 96.0 seconds	3.12 → 3.32 GiB

Combined elapsed time fell from 252.8 to 172.4 seconds: 80.4 seconds saved (31.8%) for these two files. Memory was not uniformly lower: federal-credit peak RSS rose approximately 6.4% in this pair.

These are single before/after local measurements on macOS with Python 3.14.6, PolicyEngine Core 3.32.8, the frozen dependency lock, CI's itemization setting, and Hypothesis seed 20261007. Each file ran in a separate sequential process; timings include import, setup, and shutdown. Measurements used baseline a451dd5737f; the model, dependency versions, and affected test baselines are unchanged in current main c6a9b818350. Full Rest group memory and Linux CI savings remain unmeasured.

Test plan

• Policy YAML coverage: not applicable; test setup changes only.
• Both affected Python files: all 29 existing cases pass before and after, with identical collected case IDs.
• Static comparison preserves test names, parameterization decorators, and assertion counts; reviewed the setup and calculation-order changes.
• make format and git diff --check pass on the latest main base, which includes Skip the pull request workflow for draft PRs #9993.
• Changelog fragment in changelog.d/.

axiom: n/a: test setup refactor only; no policy changes.