Preserve model EOS and chat-template defaults

[Qubitium]

Summary

• preserve the checkpoint's full generation_config.eos_token_id value instead of replacing it with the tokenizer's single primary EOS id
• copy the model's loaded generation config for continuous generation and pass its default_chat_template_kwargs into chat rendering
• stop every GSM8K variant at </assistant> and cover the assistant-boundary contract with regressions
• bump Evalution to 0.0.9 and document the release in the README

Root cause

Laguna S 2.1 declares EOS ids [2, 24]; token 24 is </assistant>, while the tokenizer's primary eos_token_id is only 2. Evalution replaced the model value with the tokenizer value and rebuilt continuous-generation configuration from model.config, discarding both the assistant-boundary EOS and default_chat_template_kwargs={"enable_thinking": true}. GSM8K also lacked the textual assistant boundary in its stop list.

Tokenicer already loads the checkpoint correctly: </assistant> encodes to [24], and its clean main suite passes. The reusable fix therefore belongs in Evalution's transformer generation contract rather than Tokenicer.

Validation

• CUDA_VISIBLE_DEVICES=2,3 PYTHON_GIL=0 pytest -q -m 'not integration' — 741 passed, 274 deselected
• PYTHON_GIL=0 pytest -q tests/test_transformer.py tests/test_gsm8k.py tests/test_gsm8k_platinum.py — 86 passed
• focused regression selection — 3 passed
• ruff check on all changed files — passed
• git diff --check — passed
• pytest -q tests/test_package.py tests/test_startup_banner.py — 10 passed
• python -m build --wheel — built evalution-0.0.9-py3-none-any.whl with packaged version 0.0.9
• Tokenicer main (v0.0.14) full suite — 468 passed, 4 skipped
• live Laguna S 2.1 GPTQModel checkpoint load/request-preparation smoke on physical GPUs 2 and 3 with CPU spill — EOS [2, 24], assistant token [24], GSM8K stop contains </assistant>, prompt ends in <assistant><think>

Companion GPTQModel loader/evaluation fix: ModelCloud/GPTQModel#2961