[ChrisW09]

Fixes #529

Problem

The loaders keep a trailing partial batch, so pretraining sees tiny batches whenever the number of training or validation rows leaves a small remainder for batch_size. Such a batch can't form any pair:

• one row, regression: get_knn() squeezed the (1, 1) distance matrix to shape (1,) and topk(...).indices[:, 1:] raised IndexError;
• one or two rows, classification: no row has both a same-class and a different-class neighbor, so the "no usable sample in this batch" ValueError stopped pretraining.

Fix

ContrastivePretrainer._cannot_form_pairs(batch) is true for batches too small for any row to be an anchor: fewer than two rows for regression, fewer than three for classification. training_step, validation_step and test_step return None for such a batch, so Lightning skips it. Larger batches are handled exactly as before, including the documented ValueError when no row of a batch qualifies (e.g. three rows of three different classes).

#577 (PR #578) independently makes the training loader drop a one-row trailing training batch. This PR also covers one-row validation batches and two-row classification batches, which that change doesn't touch.

Tests

In tests/training/test_training_pretraining.py:

• test_batch_too_small_to_form_pairs_is_skipped[regression-one-row|classification-one-row|classification-two-classes|classification-one-class]: the three steps return None.
• test_three_row_batch_without_same_class_neighbors_still_raises: the documented error is kept for larger batches.
• test_estimator_pretraining_completes_with_a_tiny_last_batch[regression-one-row-val-batch|classification-one-row-val-batch|classification-two-row-train-batch]: FTTransformer*.pretrain() with batch_size=16 and 17 validation rows or 34 training rows.

All new cases except the three-row guard fail on main; all pass with this change.

Docs

The pretraining note in docs/api/training/index.rst says that batches too small to form any pair are skipped rather than raising.

Checks

• just test: 3244 passed, 46 skipped, 28 xfailed; coverage 91.30%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)