Persist the fitted estimator so an expensive ridge fit can predict in a later session

Follow-up to #531. `BrainData.write('.h5')` on a fitted object stores the `FitResult` maps, design and kind, so a reload restores `data.model` and effect-only contrasts and the no-argument `predict()` work; `compute_contrasts(inference=True)`, `predict(X=...)` and `bootstrap` raise on a reloaded record with the one-line refit, and a ridge record's `cv` is not stored. The point where the estimator would be written is marked with a comment in `nltools/io/h5.py`.

Wanted, not blocking: optionally persist the fitted estimator (Himalaya ridge weights and selected alphas are plain arrays; nilearn's GLM regression results are heavier and noise-model dependent) so an expensive fit does not have to be re-estimated for prediction in a future session. Decide the storage format per model kind, what `write()` on the record should emit for it, and how a partial reload (maps without estimator) is reported to the user.

Start in nltools/io/h5.py at the comment marking where the estimator would be written, then trace BrainData.write('.h5') and the existing FitResult reload path. Compare the ridge and nilearn GLM model kinds and determine what each should persist, including behavior when maps exist without an estimator. Done means a reloaded fitted record can use the supported prediction paths without unnecessary refitting and reports partial reloads clearly.

Written by the indexing model from the issue text.