fix: support gpt-5 chat through SkillClaw proxy

[dfitz1138]

Translate legacy max_tokens to max_completion_tokens for gpt-5/o-series chat-completions requests while preserving legacy chat model behavior.

Also keep Hermes proxy compatibility guards for unsupported reasoning_effort fields and oversized completion-token requests.

Tests: uv run pytest tests/test_hermes_proxy_compat.py tests/test_responses_native.py -q

[Upper9527]

There is a useful minimal fix here—translating max_tokens to max_completion_tokens for OpenAI GPT-5/o-series chat models—but the PR also makes two global behavior changes that are too broad:

1. reasoning_effort is stripped for every upstream, including reasoning-capable models that support it.
2. Completion tokens are hard-capped at 8192 for every compatible provider/model, even when the configured upstream supports more.

Please split this down to the model-specific token-field translation, or make stripping/capping provider-aware and configurable with tests for reasoning-capable upstreams. As written it would silently disable requested reasoning and truncate valid larger outputs.