Impute selected_marketplace_plan_benchmark_ratio from CPS premiums

[MaxGhenis]

Closes #800.

Problem

`selected_marketplace_plan_benchmark_ratio` in policyengine-us is currently an input variable with `default_value = 1.0`. Nothing populates it, so PolicyEngine-US effectively assumes every Marketplace enrollee picks the benchmark silver plan. Real CMS data shows ~35% pick bronze and ~15% pick gold/platinum — the rules-based `marketplace_net_premium` (PolicyEngine/policyengine-us#8105) overstates OOP for bronze pickers and understates for gold pickers.

Approach — CPS premium back-out

For each Marketplace-enrolled tax unit in CPS, back out the implied plan-to-SLCSP ratio from the accounting identity:

```
reported_net_premium ≈ selected_plan_cost − APTC
selected_plan_cost = SLCSP × benchmark_ratio
→ benchmark_ratio = (reported_net_premium + computed_PTC) / SLCSP
```

CPS-reported premium is net of APTC for subsidized Marketplace takers, so adding back PolicyEngine's computed PTC recovers the sticker price. Dividing by SLCSP gives the ratio.

Ratios clipped to `[0.5, 1.5]` to handle reporting noise and Marketplace-flag false positives.

Changes

• New pure-function helper `compute_marketplace_plan_benchmark_ratio` in `policyengine_us_data/datasets/cps/cps.py` — all the math, no Microsimulation, unit-testable.
• New CPS-stage function `add_marketplace_plan_benchmark_ratio` that loads the dataset, runs a Microsimulation to pull SLCSP / ACA PTC / reported premium / takeup, and writes the ratio back.
• Called in `generate()` immediately after `add_takeup` and before downsampling.
• 6 unit tests covering silver / bronze / gold / non-taker / zero-SLCSP / clipping cases — synthetic inputs only, no H5 dependency.

Relationship to #618

#618 (open, by @daphnehanse11) adds state-level calibration targets for total APTC-taking Marketplace tax units and the bronze-plan subset, from CMS 2024 OEP data. That constrains the aggregate weighted distribution.

This PR is complementary — it sets the per-household ratio so each household's computed `marketplace_net_premium` reflects the plan they actually picked. After both land, the state-level targets from #618 can validate that the aggregated ratio distribution lines up with CMS effectuated-plan metal shares.

Caveats

• Marketplace flag noise: CPS respondents sometimes confuse Marketplace with employer / Medicaid coverage. The `takes_up_aca_if_eligible` gate filters most, and the [0.5, 1.5] clip catches the rest.
• CSR silver variants: cost-sharing-reduction silver plans have the same sticker premium but reduced deductibles/copays. The back-out correctly lands on ratio ≈ 1.0; plan-value differences are a separate MOOP question.
• Non-takers / zero SLCSP: keep the 1.0 default. These paths are zeroed out downstream by `takes_up_aca_if_eligible` gating inside `selected_marketplace_plan_premium_proxy`, so the default never leaks into calculated OOP.

Testing

`uv run pytest tests/unit/datasets/test_marketplace_plan_benchmark_ratio.py -v` → 6/6 pass.

Test plan

• Unit tests pass
• `ruff check` clean on new code (pre-existing unrelated errors in cps.py are unchanged)
• CI passes
• Integration: verify weighted ratio distribution across Enhanced CPS approximately matches CMS metal-tier shares (manual check after data build)