Fix Laguna Evalution runtime configuration

[Qubitium]

Summary

• treat legacy random_seed as an Evalution engine option and normalize it to Evalution's canonical seed
• stop the model-test loader from passing both seed and random_seed, preventing random_seed from leaking into LagunaForCausalLM.__init__
• run Laguna S 2.1 GSM8K Platinum through the chat-template path so the checkpoint's required thinking default is active

Root cause

The Evalution adapter recognized seed but not random_seed, so the latter remained in Model.model_kwargs and was forwarded through GPTQModel.load() to the Hugging Face model constructor. Separately, Laguna S 2.1's evaluation config used a plain prompt even though its reasoning mode is selected by the chat template and declared as default_chat_template_kwargs={"enable_thinking": true} in generation_config.json.

This change keeps seed canonical, accepts random_seed only as a backwards-compatible engine alias, and enables chat-template evaluation specifically for Laguna S 2.1.

Validation

• CUDA_VISIBLE_DEVICES=2,3 PYTHON_GIL=0 pytest -q tests/test_eval_loader_args.py tests/test_model_test_helpers.py tests/test_laguna_support.py — 18 passed
• ruff check on all changed files — passed
• git diff --check — passed
• live Laguna S 2.1 GPTQModel checkpoint load/request-preparation smoke on physical GPUs 2 and 3 with CPU spill — engine seed 38, empty model kwargs, EOS [2, 24], </assistant> token [24], and prompt suffix <assistant><think>

Companion Evalution fix: ModelCloud/Evalution#128