Implement core ML bridging utilities

[ysims]

Replace the NotImplementedError placeholders in earthrs.ml.core with real, optional-dependency-guarded implementations:

• to_numpy: converts Samples, row-mapping lists, or column-oriented mappings to (array, column_names); plain array-likes convert to a bare array. Falls back to dtype=object when values aren't numeric so nothing is lost.
• prepare_training_data: splits a Samples table into aligned (X, y) via to_numpy, accepting either a label column name or a separate aligned label sequence, with optional NA-row dropping (None/NaN, matching the None-for-missing convention already used in earthrs.processing.core).
• to_torch_dataset / to_tensorflow_dataset: wrap features/labels in a TensorDataset / tf.data.Dataset via numpy.
• tile_scene: splits a Scene's band grids into tile_size x tile_size windows (stride defaults to tile_size), cropping partial edge tiles, offsetting scene.transform per tile, and preserving crs/sensor/ acquisition_time/history plus a tile_offset in metadata.
• raster_predict: tiles a scene, stacks each tile's bands into a numpy array, batches tiles, and calls model.predict/model(), returning (tile, prediction) pairs so callers can reconstruct the full raster.

Add ml / ml-torch / ml-tensorflow optional-dependency extras to pyproject.toml, and tests/test_ml.py covering all six functions (numpy/ torch/tensorflow paths use pytest.importorskip so the suite stays green without those installed).