[Migration] Regenerate the recorded trace fixtures that predate bambi 0.20 (*_Intercept_centered)

Nobody has claimed this yet.

Part of #1305. Parent: #1306.

⚠️ **Found while fixing F6 (#1315)** — it was masked by R10. Removing the R10 xfail marks on the 16 `tests/test_plotting_cartoon.py` ids made them fail one step later, and the same gap explains why the R2-marked `test_sample_posterior_predictive` grid still fails after the prediction path was fixed.

From bambi 0.20's *Intercept handling* (see F12, #1330): every common intercept is now a `*_Intercept_centered` free RV with the uncentered `*_Intercept` kept as a deterministic. HSSM keeps the centered RV in `posterior` because bambi's `predict` and `compute_log_likelihood` (via `pm.compute_deterministics`) raise `KeyError` without it.

The recorded trace fixtures under `tests/fixtures/` predate that change and only carry `*_Intercept`:

| fixture | loaded by | missing |
|---|---|---|
| `cavanagh_idata.nc` | `cav_dt` (`tests/conftest.py`) | `v_Intercept_centered` |
| `idata_cavanagh_cartoon.nc` | `cav_model_cartoon` | `v_Intercept_centered`, `a_Intercept_centered` |
| `test_idata_race.nc` | `race_model_cartoon` | `v0_Intercept_centered`, `a_Intercept_centered` |
| `idata_race_cartoon.nc` | (no test references it) | `v0_Intercept_centered`, `a_Intercept_centered` |

```
KeyError: 'v_Intercept_centered'
src/hssm/plotting/model_cartoon.py:397 compute_merge_necessary_deterministics
pymc/sampling/deterministic.py:176 compute_deterministics
```

Two consequences today:

- Any in-sample `Model.predict` / `compute_deterministics` on these traces raises `KeyError` (the cartoon tests, the `test_sample_posterior_predictive` grid).
- Out-of-sample `predict(data=...)` on them does **not** raise: `pm.sample_posterior_predictive` silently resamples the missing RV from its prior, so the quantile-probability and predictive plotting tests pass on draws that are partly prior, not posterior.

### Tasks
- [ ] Regenerate the trace fixtures with the bambi 0.20 model definitions used by their fixtures (`cav_model_cartoon`, `race_model_cartoon`, the `v ~ theta + (1|participant_id)` model behind `cav_dt`), keeping draw counts small; decide whether `idata_race_cartoon.nc` is dead and can be deleted
- [ ] Add a regeneration script (or document the recipe) so the fixtures can be refreshed on the next bambi change instead of being re-discovered
- [ ] Remove the stale-fixture xfail marks in `tests/test_plotting_cartoon.py` (16 ids) and re-check the R2 marks on `tests/test_sample_posterior_predictive.py::test_sample_posterior_predictive`
- [ ] Consider a guard: fail loudly (or warn) when `sample_posterior_predictive` is handed a trace missing free RVs of the current model, so prior resampling cannot pass as posterior prediction

Details: `design/bambi-migration-fix-plan.md` (F6 checklist), `design/bambi-dev-upgrade-test-failures.md` (R10 status note).

Start with the fixture loaders in tests/conftest.py, the model definitions for cav_model_cartoon and race_model_cartoon, and the fixtures under tests/fixtures/. Run the affected cartoon tests and tests/test_sample_posterior_predictive.py::test_sample_posterior_predictive to reproduce the failures. Done means regenerated fixtures, refreshed xfail decisions, a documented regeneration recipe, and a resolved decision on idata_race_cartoon.nc and the missing-RV guard.

Written by the indexing model from the issue text.