fix(python): forward unk_id in SentencePieceUnigramTokenizer

[Con-Benksl]

Fixes #1576.

Even with <unk> in a custom vocabulary, SentencePieceUnigramTokenizer cannot encode unknown characters because its constructor does not expose the underlying Unigram model's unk_id. Add an optional trailing parameter and forward it to the existing model constructor, preserving the default behavior and positional arguments. Also pass unk_id by keyword in the parity checker, where it was incorrectly bound to replacement.

Validation on Python 3.13.13 with the native extension built from this checkout:

• PYTHONPATH=py_src HF_HUB_OFFLINE=1 HF_DATASETS_OFFLINE=1 python -m pytest -q tests/implementations/test_sentencepiece.py: 7 passed; before the fix, the two new cases failed and the five existing cases passed. The regression covers zero/nonzero IDs, default behavior, positional compatibility, and missing-vocabulary validation.
• Ran the actual parity checker against a tokenizer trained locally on three short sentences, with unk_id=2 and an unseen character in the input: all three encodings matched SentencePiece. The original call failed with a TypeError.
• make check-style passed; final ruff checks, formatting checks, and the repository's ty check also passed after excluding unrelated generated stub changes from the diff.

No model files were downloaded. The full network-dependent test suite was not run.

AI assistance: OpenAI Codex implemented and tested this change; a separate Codex agent reviewed the diff.