Add training and export functionality with Python bindings

[mosuka]

This commit introduces model training and dictionary export capabilities
to lindera-python, enabling users to train custom morphological analysis
models from annotated corpus data.

Features:

• Add train() function to train CRF-based models from corpus
  • Supports L1 regularization, configurable iterations, and multi-threading
  • Accepts seed lexicon, corpus, character/unknown word/feature definitions
• Add export() function to export trained models to dictionary files
  • Generates lex.csv, matrix.def, unk.def, char.def
  • Optional metadata.json update support

Implementation:

• New src/trainer.rs module with PyO3 bindings for train/export
• Add 'train' feature flag in Cargo.toml (requires lindera/train)
• Use local lindera path (../lindera/lindera) for latest trainer API
• Add num_cpus dependency for automatic thread detection

Documentation:

• Update README.md with training/export usage examples
• Add examples/train_and_export.py with complete workflow demonstration
• Add tests/test_trainer.py with comprehensive test coverage
• Corpus format follows lindera/resources/training conventions
  (tab-separated surface + features with EOS markers)

Changes:

• Modified: Cargo.toml, src/lib.rs, README.md
• Added: src/trainer.rs, examples/train_and_export.py, tests/test_trainer.py
• Updated: Cargo.lock, poetry.lock, pyproject.toml, Makefile