Add sklearn-compatible adapter (anomaly, forecasting, embedding)

[DrLHS]

This PR adds momentfm.sklearn_adapter, three sklearn estimator-API adapters that let MOMENT drop into sklearn.pipeline.Pipeline, ColumnTransformer, GridSearchCV, and any other tooling that follows the sklearn contract.

Adapters

• MOMENTAnomalyDetector(TransformerMixin) — reconstruction-MSE per-sample anomaly score. transform(X) returns shape (n_samples,); score_samples(X) returns the negated score in the sklearn outlier-detector convention (higher = more normal).
• MOMENTForecaster(RegressorMixin) — predict(X) returns the forecast horizon as a flat 2D array of shape (n_samples, n_channels * forecast_horizon) so it slots into sklearn pipelines without a custom reshape step. Construct with MOMENTForecaster(forecast_horizon=H).
• MOMENTEmbedder(TransformerMixin) — transform(X) returns the encoder embedding as a 2D feature matrix (n_samples, embedding_dim) so downstream sklearn estimators (LogisticRegression, KMeans, t-SNE) can consume them directly.

Input shape

All three adapters accept either 2D (n_samples, n_channels * context_length) input (sklearn pipeline convention) or 3D (n_samples, n_channels, context_length) (MOMENT native). Shape coercion is centralised in a _to_3d helper with explicit error messages.

fit contract

fit is zero-shot by default — it loads the pretrained weights and sets the model to eval(). This matches MOMENT's intended zero-shot-foundation-model usage. Calling predict or transform before fit is equivalent to calling fit first.

Tests

tests/test_sklearn_adapter.py adds 18 tests:

• 6 shape-coercion tests (2D passthrough, 2D → 3D, multichannel, rejection of 4D, rejection of wrong 2D width, rejection of wrong 3D shape)
• 6 sklearn API conformance tests (`get_params` and `set_params` parametrised across all three adapters)
• 1 `MOMENTForecaster` validation test (must set `forecast_horizon` before `fit`)
• 5 integration tests covering the real forward pass for each adapter plus a sklearn-Pipeline integration test. Marked `skip` by default so CI doesn't need network access; flip the `skipif` to `False` to enable when weights are cached locally.

13 unit tests pass on the current branch; 5 integration tests correctly skip.

README

A new ## 🧩 scikit-learn Integration section just before the tutorials section shows the three usage idioms.

What this enables

• Drop-in MOMENT for anyone with an existing sklearn-pipeline-based feature engineering or model stack
• Reproducible grid search over MOMENT model sizes via standard sklearn `GridSearchCV`
• Mixed pipelines that pre-process tabular features upstream and feed MOMENT downstream
• A natural integration target for tabular ML practitioners who don't want to write a custom PyTorch loop

Files changed

• `momentfm/sklearn_adapter.py` (new, ~250 lines)
• `tests/test_sklearn_adapter.py` (new, 18 tests)
• `README.md` (one new section, 35 lines)

Companion to #85 (CI) and #84 (release-cut request).