v0.11.0 — Self-Supervised Pretraining (Masked Language Modeling)

Self-Supervised Pretraining (Masked Language Modeling). Pretrain Polaris' own
transformer on unlabeled text with an MLM objective, then transfer the trunk into
a classifier and fine-tune — the BERT recipe, from scratch, no downloads. A
controlled ablation (same vocabulary and 4-layer architecture, pretraining the only
difference) shows pretraining works in the expected direction — a large warm start
(epoch-1 validation 0.810 vs 0.736), faster convergence, higher best validation
(0.864 vs 0.851) — but converges to the same ~86% test ceiling (0.853 vs 0.852),
because 25k labels already suffice and the small in-domain corpus injects little new
knowledge. Four from-scratch levers (transformer, BPE, GloVe, MLM pretraining) now
all land at ~85-86%: the ceiling is the task and data/compute regime, not any one
component.

Added

• polaris.pretraining: masked-language-model pretraining from scratch —
  mask_tokens / MaskedLMBatch (BERT-style 80/10/10 masking with dynamic,
  seedable corruption), MaskedLanguageModel (shared trunk + vocabulary head,
  with transfer_encoder_to for moving a pretrained trunk into a classifier),
  and pretrain (the MLM loop: masked-position cross-entropy, warmup scheduling,
  per-epoch loss and masked-token accuracy).
• TransformerEncoder (polaris.models): the shared, headless transformer trunk
  (embedding + positional + encoder blocks + final norm → per-token hidden
  states), now reused by both the classifier and the masked-language model.
• mask_token / mask_id on Vocabulary, and a mask_token argument on
  build_vocabulary.
• IMDB's "unsupervised" split (50,000 unlabeled reviews) is now loadable via
  IMDBDataset, for pretraining.
• examples/pretrain_finetune_imdb.py: the full pretrain → transfer → fine-tune
  thread on IMDB.

Changed

• TransformerEncoderClassifier is refactored to compose the shared
  TransformerEncoder trunk (behavior-preserving; its trunk weights now live
  under the encoder submodule).