Normalize Mistral tokenizer regexes for Laguna

[Qubitium]

Summary

• enable the Transformers fix_mistral_regex correction by default for Tokenicer string loads
• preserve explicit caller overrides and carry normalized kwargs through fallback loaders
• retry tokenizer-only loading for the newer Hugging Face strict dataclass base exception
• add regression tests and bump Tokenicer to 0.0.14

Root cause

Laguna S 2.1 ships a Mistral-family pre-tokenizer regex that Transformers identifies as incorrect. GPTQModel and other downstream loaders should receive the correction from Tokenicer instead of maintaining per-model tokenizer patches.

Validation

• python -m pytest -q tests/test_mistral_regex.py — 3 passed
• python -m pytest -n 8 -q tests — 467 passed, 4 environment/checkpoint skips
• real /monster/data/model/Laguna-S-2.1 comparison — corrected token boundaries for eBay, mL, NaOH, and pH
• python -m ruff check tokenicer/tokenicer.py tests/test_mistral_regex.py tests/test_loop_models.py
• source distribution build for 0.0.14 and twine check