fix: remove train/evaluate imports from __init__, pin tokenizers==0.19.1

fix: default MODEL_NAMES to fine-tuned HF repos, fix .gitignore

feat: Phase 1 enhancements - ensemble endpoint, history API, rate limiting, storage monitoring

fix: smart MODELS_DIR path resolution for Docker

fix: remove deprecated use_cache from xlnet config

chore: clean up preload log message

fix: patch xlnet tokenizer_config at runtime

fix: preload models synchronously

fix: sync latest inference and api fixes

fix: Python 3.9 compat list[dict] type hints

fix: preload all models at startup

fix: evict model cache to prevent OOM

fix: remove syntax error in xlnet inference

fix: xlnet list fields in _token_importance

fix: xlnet tokenizer list fields

fix: distilbert token_type_ids and xlnet inference errors

Initial deployment