Add ChatGPT OAuth model backend and fix evolution runtime regressions

[kyzer1022]

Summary

• add a ChatGPT OAuth-backed DSPy LM for chatgpt/* models so skill evolution can run without OpenAI API keys
• rebuild ChatGPTOAuthLM instances on copy() so DSPy avoids deepcopying the OpenAI/httpx client stack during optimization
• fix evolve_skill.py to prefer explicit dataset_path loading, validate the full skill text, and pass valset into the MIPRO fallback path so tiny datasets stop crashing
• add regression coverage for OAuth token resolution/request building, LM copy behavior, dataset-path precedence, full-skill validation, and GEPA → MIPRO fallback wiring

Testing

• .venv/bin/pytest -q
• .venv/bin/python -m evolution.skills.evolve_skill --skill github-code-review --iterations 1 --dataset-path /home/kyzer/Documents/projects/hermes-agent-self-evolution/tmp/github-code-review-mini --optimizer-model chatgpt/gpt-5.4 --eval-model chatgpt/gpt-5.4 --hermes-repo /home/kyzer/.hermes/hermes-agent

Notes

• Verified that a real evolution run now completes end-to-end with ChatGPT OAuth models.
• The latest mini-dataset run completed successfully and saved artifacts under output/github-code-review/20260414_164858/.
• The generated baseline/evolved skill artifacts were still byte-identical, which points to a separate optimizer/artifact-mutation issue outside the runtime fixes in this PR.