Migrate all mini-batch (_many) methods to narwhals for dataframe-agnostic support

Motivation

#1900 introduced dataframe-agnostic mini-batching via narwhals: inputs are wrapped at the method boundary, the numpy compute core stays untouched, and outputs are rebuilt in the caller's native backend (pandas / polars / pyarrow / nullable / arrow-backed pandas). The helpers live in river/utils/dataframe.py:

• into_frame / into_series — wrap native inputs
• to_numpy — extract a float64 array for the compute core
• to_native_frame / to_native_series — rebuild output in the caller's backend, preserving the pandas index

So far only linear_model (GLM + LinearRegression, LogisticRegression, BayesianLinearRegression) uses this. Every other mini-batch method still hard-codes pandas (pd.DataFrame/pd.Series signatures, .values, .columns, etc.). The goal of this issue is to migrate all _many methods to narwhals so that any narwhals-supported backend works end to end.

Scope

Mini-batch methods to migrate (learn_many, predict_many, predict_proba_many, transform_many, and friends):

Base classes (signatures/type hints — do first, they're authoritative):

• base/classifier.py — MiniBatchClassifier
• base/regressor.py — MiniBatchRegressor
• base/transformer.py — MiniBatchTransformer

Concrete estimators (still pandas-only):

• preprocessing/scale.py — StandardScaler & co. (.values/.columns) https://github.com/online-ml/river/pull/1932
• naive_bayes/base.py — BaseNB (joint_log_likelihood_many, predict_proba_many, pandas .subtract)
• multiclass/ovr.py — OneVsRest https://github.com/online-ml/river/pull/1933
• anomaly/lof.py — LocalOutlierFactor https://github.com/online-ml/river/pull/1929
• anomaly/svm.py — OneClassSVM https://github.com/online-ml/river/pull/1930
• feature_extraction/vectorize.py https://github.com/online-ml/river/pull/1904
• preprocessing/ordinal.py — OrdinalEncoder https://github.com/online-ml/river/pull/1921
• preprocessing/one_hot.py — OneHotEncoder https://github.com/online-ml/river/pull/1926
• covariance/emp.py — EmpiricalCovariance (and EmpiricalPrecision) https://github.com/online-ml/river/pull/1923

Composition (should mostly fall out once the above are done, but need verifying):

• compose/pipeline.py — Pipeline https://github.com/online-ml/river/pull/1932
• compose/union.py — TransformerUnion https://github.com/online-ml/river/pull/1932
• compose/func.py — FuncTransformer https://github.com/online-ml/river/pull/1932
• compose/product.py — TransformerProduct https://github.com/online-ml/river/pull/1932
• compose/select.py — Select https://github.com/online-ml/river/pull/1932

Approach

For each method:

1. Wrap inputs with into_frame / into_series at entry.
2. Drop to numpy via to_numpy for the compute core (leave the math unchanged).
3. Rebuild outputs with to_native_frame / to_native_series so the caller's backend and index are preserved.
4. Replace pd.DataFrame / pd.Series type hints with the narwhals IntoDataFrame / IntoSeries aliases.
5. Add multi-backend tests using the existing frame_backend fixture (river/conftest.py), mirroring test_glm.py.

learn_many/predict_many outputs must stay byte-for-byte identical for the pandas path so nothing regresses.

Notes / gotchas (from #1900)

• to_numpy forces float64 coercion because pandas ArrowDtype columns otherwise come back as object and break downstream ufuncs.
• Non-pandas backends require string column labels; to_native_frame stringifies for those.
• The pandas index is only carried over when nw.maybe_get_index() is not None.