fix(providers): fall back to Responses API for compatible endpoints

[aniketshukla1]

Summary

• Keep custom endpoints on Chat Completions by default and switch only after an endpoint explicitly rejects a tool-enabled request and directs it to /v1/responses.
• Reuse the existing OpenAI SDK client and cache a working Responses protocol per model.
• Preserve Responses tool-call IDs and avoid retrying a stream after Chat Completions output has started.

Fixes #85

Validation

• .venv/bin/pytest -q (1116 passed, 1 skipped)
• git diff --check

Screenshots

Not applicable: this is provider protocol routing and regression coverage with no UI change.