fix: gate reasoning effort and support reasoning content in streams ✨ (#93)

**🐛 Fixes**
- Fixes #92
- Hide the `reasoning_effort` picker unless the model explicitly supports that parameter, and omit `"none"` from requests entirely. This prevents rejected calls and retry loops on backends that do not accept the field.
- Treat `reasoning_content` in streamed chat deltas as thinking output by default, with an option to merge it back into `content` for compatibility. This improves support for models like Qwen3 and llama.cpp.
- Expand parameter support checks and request building so reasoning settings are only sent when allowed by the model metadata.

**🧪 Tests**
- Add coverage for `reasoning_effort` support detection, omission of `"none"`, and request serialization behavior.
- Add streaming tests for `delta.reasoning_content` handling with both separate-thinking and merged-content modes.

**🧹 Chores**
- Bump the extension to `2.1.1`, mark the package as preview, and update the changelog.
- Add the feature-request prompt and refresh the investigative agent tool list.