🚀 Release v2.1.3 — Image content fix & stream recovery (#99)

* chore: ✨ bump dev version and expand marketplace keywords

**🧹 Chores**
- Bump the extension version to `2.1.3-dev1` to reflect the next development build.
- Add additional marketplace keywords to improve discoverability for language model and Copilot connector searches.

* fix: recover buffered tool calls on incomplete streams ⚡

**🐛 Fixes**
- Treat clean SSE closures without a `[DONE]` marker as implicit completion when no buffered data remains. This avoids false stream-end errors for long-running proxy responses.
- Detect truly truncated streams by checking for leftover buffer data or empty payload streams, and emit clearer structured logs for each case.
- Add `flushPendingBuffers()` to emit buffered `/responses`, anonymous, and OpenAI tool calls when a stream ends early. The provider now attempts a best-effort recovery instead of failing hard.
- Wrap `image_url` content items in arrays for LiteLLM `/responses` requests. This prevents `ValueError: Invalid content type: <class 'dict'>` for both user and assistant vision messages.

**🧪 Tests**
- Add regression coverage for clean SSE termination, truncated buffers, and buffered tool-call flushing.
- Add adapter tests to verify image content is array-wrapped and preserve mixed-content behavior.
- Add provider-level recovery tests for incomplete streams without a `[DONE]` marker.

* test: add regression coverage for inline-edit image content in responses 🧪

**🐛 Fixes**
- Add a regression test for bug #98 to ensure inline-edit sessions with `image_url` content never serialize to bare-object message content in `/responses` requests.
- Verify image-bearing messages are array-wrapped and that function call/output items remain present and correctly linked.

* chore: update CHANGELOG for v2.1.3 📝

* chore: include package.json v2.1.3 in release commit 📦