Retry with max_completion_tokens when the endpoint rejects max_tokens

[Schro-code]

Summary

• Problem: Newer OpenAI reasoning models (o1, o3, GPT-5 family) renamed max_tokens to max_completion_tokens and return a 400 error with error.code: "unsupported_parameter" when the old name is sent. They also reject temperature values other than 1 with error.code: "unsupported_value".
• Fix: Adds a send_with_compat_retries helper that inspects the structured error.param / error.code fields in the 400 response (not free-text matching) and retries with the corrected parameter. Used by the non-streaming chat_at path.
• Streaming path: The same compat-retry logic is woven into chat_at_streaming's existing retry loop, so both code paths handle the transition transparently.
• Loop safety: Each adjustment applies at most once per request — the already-applied key is removed from the body before retrying, so the loop terminates even if the endpoint keeps rejecting.

Test plan

• Verify existing tests still pass (cargo test -p utopia-llm)
• Manual test against an o1/o3/GPT-5 endpoint: confirm the 400 → retry → success flow for max_tokens rename
• Manual test: confirm temperature rejection is handled gracefully (parameter dropped, retry succeeds)
• Confirm non-reasoning models (e.g. GPT-4o) are unaffected (no retry needed, passes through directly)

🤖 Generated with Claude Code

[Schro-code]

Closing to reopen with correct base branch (dev) and signed-off commit per CONTRIBUTING.md.