Weight validation minibatch losses by their example counts

[sylvesterkaczmarek]

Summary

Fixes #52.

Aggregate validation cross entropy by example count instead of giving every minibatch equal weight. A partial final minibatch previously had disproportionate influence on validation scores and best-checkpoint selection. Training updates, training-loss reporting, regularization and the TPR-at-FPR branch are unchanged.

Validation

python -m pytest -q src/synthid_text/logits_processing_test.py src/synthid_text/detector_validation_test.py

23 tests passed. All existing logits-processing tests and eight new validation cases pass. Five new cases fail on unchanged upstream; three partition controls pass. The tests execute the actual Flax detector and train loop, compare every epoch with direct full-set cross entropy, test validation-order invariance, and verify checkpoint selection after a real two-epoch parameter update. The existing suite emitted five PyTorch batching-performance warnings. No language-model checkpoint, tokenizer download, generated corpus, GPU or production detector calibration was used. This does not change single-epoch/no-validation edge cases or claim CPU training quality at production scale.

Tested locally on macOS CPU using real module imports. New-test formatting, scoped static checks, syntax checks and git diff --check pass. No dependency or workflow changes. The full repository suite was not run.