🚀 Release v2.2.1 reasoning fixes (#119)

* fix: refine reasoning-effort fallback for LiteLLM models

**🐛 Fixes**
- Keep the default reasoning ladder at `none/low/medium/high` and merge in only explicitly true LiteLLM effort fields when `supports_reasoning` is enabled. This exposes extended levels like `xhigh` and `max` without broadening unsupported options.
- Treat all-false LiteLLM effort metadata as authoritative and avoid inferring support from null fields. This prevents the UI from showing reasoning choices for models that explicitly disabled them.

* fix: tighten reasoning capability detection and make model overrides opt-in

**🐛 Fixes**
- Apply matching model-card overrides before reasoning capability checks so only explicitly named LiteLLM fields are replaced. Unspecified capability fields stay unchanged, and null values are preserved.
- Hide reasoning controls when LiteLLM explicitly disables reasoning or reports only partial effort metadata. Explicit effort levels are now used as-is instead of inferring sibling levels.
- Default `litellm-connector.enableModelOverrides` to `false` so override patches are opt-in.

**🧪 Tests**
- Add regression coverage for disabled overrides, field-level override patching, and explicit reasoning effort handling. This protects the new reasoning gates and merge behavior.

**📚 Docs**
- Update the README and marketplace notes with the opt-in override workflow and revised reasoning guidance.
- Refresh the changelog and package metadata for the new dev release.

* 🚀 Release v2.2.1 reasoning fixes