feat: add an estimator interface delegating to train

[andremsouza]

fit, transform, predict, fit_transform, score, get_params, set_params on SOM. Entirely
additive: train is unchanged and remains the primary way to train. Second of three PRs for
0.6.0.

Modelled on KMeans, verified against the installed 1.9

KMeans	SOM
fit(X, y=None)	same	returns the estimator
transform(X) → (n, n_clusters)	(n, x*y)	distance to every node
predict(X) → (n,)	(n,)	flat node index
score(X)	same	negated, so larger is better
cluster_centers_	weights_
inertia_	quantization_error_	None before training
n_features_in_	same

On Ousterhout

These are pass-through methods by the definition in A Philosophy of Software Design §7.1 — "does
little except invoke another method, whose signature is similar or identical" — and that is ordinarily
a smell. They are here because their value is conformance to an external convention rather than
abstraction: Pipeline, GridSearchCV and cross_val_score call these exact names.

The mitigation is to keep them one-liners and prove they cannot drift:

@pytest.mark.parametrize("mode", list(TrainingMode))
def test_fit_and_train_produce_identical_weights(mode):
    ...
    assert np.abs(fitted.get_weights() - trained.get_weights()).max() == 0.0

Exactly 0.0, for every mode. A pass-through that has diverged from what it passes through is the
failure this guards.

Two decisions worth reviewing

predict returns a flat index, not (row, column). A 1-D label array is what scorers,
confusion_matrix and cross_val_score assume, so coordinates would read better for a grid and
compose with nothing. np.unravel_index(som.predict(X), som.get_shape()) recovers the position, and
a test asserts it matches winner() sample by sample. winner() itself is unchanged.

score is the negated error. Every scikit-learn scorer treats larger as better and
GridSearchCV maximises. Without the sign, a parameter search would confidently select the worst
map. A test asserts a trained map outscores an untrained one, so the sign is checked rather than
assumed.

Two differences from scikit-learn, deliberate and documented

• fit continues rather than resetting. scikit-learn estimators conventionally discard fitted
  state on refit. A SOM's models are its state, and train has always continued. Pinned by a test
  so it stays a decision rather than becoming a surprise.
• The models exist before training, because initialization is a separate step. There is no
  not-fitted state to raise about.

set_params refuses some things

It changes the rates, radii, decays and distance function. It raises for the grid shape and
input_len:

ValueError: 'x' cannot be changed after construction: the models would no longer match it.

Silently rebuilding would discard trained weights; leaving them would give a map whose models
contradict its own description. It also re-runs learning-rate validation, so it cannot be used to get
around the constructor's check.

set_params is what set_learning_rate/set_neighborhood_radius become. Both still work, and a test
asserts the two routes agree.

Not sufficient for Pipeline on its own

Stated plainly because the earlier plan got this wrong: since scikit-learn 1.7, Pipeline.predict,
GridSearchCV and cross_val_score require __sklearn_tags__. These method names alone give you
fit and nothing else. The adapter that provides the rest is the next PR, keeping scikit-learn an
optional dependency.

Verification

407 tests (24 new), 100% coverage, ruff / ruff format / mypy --strict clean, mkdocs build --strict exits 0, bandit clean, architecture profile --strict passes. New docs page in the nav.