fix: support LM Studio JSON response format

[jia030928]

LM Studio rejects Memmy JSON requests because it accepts response_format.type values json_schema or text, while the OpenAI-compatible client always sent json_object. This caused both memory summary and reflection jobs to fail with HTTP 400.

The client now retries this specific compatibility error with text and remembers that capability for later JSON requests. Explicit response-format overrides remain unchanged, and unrelated 400 responses still fail normally.

Validation:

• Real LM Studio google/gemma-4-e4b: summary and reflection both succeeded
• Packaged Memory runtime against LM Studio: summary and reflection both succeeded
• npm run test -w @memmy/memory (119 files passed; 1092 tests passed; 1 skipped)
• npm run lint
• npm run typecheck
• macOS arm64 CN unsigned package built and DMG verified