regime_discovery.ipynb: detector rankings come from a single seed and don't reproduce

Section 8 of `examples/regime_discovery.ipynb` compares changepoint detectors across three regime-change types, and section 9 turns those results into a "which detector for which change type" recommendation table. Each row rests on **one dataset at one seed**, and the rankings do not survive repetition.

Re-running all three comparisons over 10 seeds, same generative process, F1 mean ± sd:

| Detector | Variance shift | Shape shift | Gradual drift |
|---|---|---|---|
| Ensemble | **1.000 ± 0.000** | 0.927 ± 0.125 | 0.725 ± 0.149 |
| PELT | 0.967 ± 0.105 | **0.986 ± 0.045** | **0.854 ± 0.091** |
| HMM | **0.986 ± 0.045** | **0.986 ± 0.045** | 0.724 ± 0.129 |
| HMM (auto) | **0.986 ± 0.045** | 0.886 ± 0.314 | 0.790 ± 0.306 |
| KS | **0.986 ± 0.045** | 0.907 ± 0.105 | 0.776 ± 0.077 |
| Levene | **0.986 ± 0.045** | 0.870 ± 0.109 | 0.714 ± 0.080 |
| CUSUM (mean) | 0.957 ± 0.069 | **0.986 ± 0.045** | 0.688 ± 0.105 |
| CUSUM (var/both) | 0.942 ± 0.125 | 0.837 ± 0.116 | 0.698 ± 0.125 |
| Baseline | 0.779 ± 0.088 | 0.665 ± 0.083 | 0.654 ± 0.196 |

Against the single-seed numbers currently in the notebook, `Ensemble` moves from near-worst to best on variance shift and `PELT` moves from joint-first to sixth. No detector wins everywhere. `HMM (auto)` is worth a look — sd above 0.30 on two of three, with a minimum of 0.000, and the run emits `Model is not converging`.

**What needs doing**

1. **Make the data reproducible.** Cells 3, 20, 24 and 27 generate data with bare `np.random.randn()`, and the only seeding is a single `np.random.seed(42)` in cell 1. Each dataset therefore depends on cumulative RNG state and changes if cells are run out of order. Give each generator its own `np.random.default_rng(seed)`.
2. **Run each comparison over several seeds** — 10 is enough to show the effect — and report **mean ± sd** rather than a single number.
3. **Handle `hmmlearn` being absent.** `HMMRegimeDetector` requires it and it is deliberately not a dependency of `causalts`. Skip the HMM rows on `ImportError`, with a note that `pip install hmmlearn` enables them, so the notebook runs on a plain install.
4. **Update section 9 so its recommendations follow from the new numbers.** The current table names a single winner per row, which error bars won't support. One correction that's independent of the sweep: PELT's stated limitation "Only mean shifts" is wrong — what PELT detects is set by its cost function, and `PeltRegimeDetector` already has a `signal="volatility"` option.

**Keep HMM.** It scores 0.986 / 0.986 / 0.724 and its worst single seed anywhere is 0.600 — it never collapses. This is about the methodology, not about demoting HMM.

The sweep used 10 seeds and F1 only, with `default_rng(seed)` in place of the notebook's global `np.random` — the same generative process, but not the same draws.

Start with examples/regime_discovery.ipynb, especially cells 1, 3, 20, 24 and 27, then inspect sections 8 and 9. Run the notebook with multiple seeds and verify that results report mean ± sd, HMM rows are skipped with an installation note when hmmlearn is absent, and the recommendations and PELT limitation match the results.

Written by the indexing model from the issue text.