Add Laguna S 2.1 model support

[Qubitium]

Summary

• support Laguna S 2.1 checkpoints with singular shared-expert naming
• keep the large-model fast quantization test to the minimum two layers that span dense and sparse/MoE topology
• configure post-quant multi-GPU Evalution coverage for ARC Challenge and GSM8K Platinum
• add a repository agent skill that routes reusable tokenizer and chat-template normalization into ModelCloud/Tokenicer
• require Tokenicer 0.0.14 and delegate GPTQModel tokenizer construction and normalization to it

Tokenizer normalization

The Laguna Mistral-regex correction is merged upstream in ModelCloud/Tokenicer#61 and released in Tokenicer 0.0.14. GPTQModel now requires tokenicer>=0.0.14, removes its duplicated tokenizer dispatch and legacy-constructor compatibility implementation, and leaves fix_mistral_regex ownership entirely in Tokenicer.

Validation

• python -m pytest -q tests/test_qwen2_family_compat.py tests/test_laguna_support.py tests/test_local_model_paths.py — 40 passed
• python -m pytest -q tests/test_hf_config_compat.py — 34 passed, 1 unrelated skip
• released Tokenicer 0.0.14 loaded /monster/data/model/Laguna-S-2.1 and round-tripped a tokenizer sample
• Tokenicer focused Mistral-regex regression — 3 passed
• complete Tokenicer suite — 467 passed, 4 environment/checkpoint skips
• GPTQModel tokenizer-normalization skill validation — passed
• Ruff, pip check, and git diff --check — passed