feat: convert CAF audio and GIF attachments for model compatibility

[mfzeidan]

iMessage voice messages use Opus-in-CAF format which isn't supported by OpenAI Whisper. GIF images aren't accepted by all model APIs (e.g. xAI). This adds ffmpeg-based conversion: .caf→.m4a for audio, .gif→.png (first frame) for images. Converted files are cached alongside originals.

[steipete]

Thanks @mfzeidan. I landed a maintainer rewrite on main in 8dcb9d0 and credited you in the commit/changelog.

The PR was directionally right, but I did not merge it as-is because current main has a batched attachment metadata path that the PR missed, it hardcoded /opt/homebrew/bin/ffmpeg, converted as a side effect of normal metadata reads, and replaced the meaning of original_path with the converted file path.

The landed version keeps conversion opt-in via --convert-attachments / RPC convert_attachments, finds ffmpeg from PATH plus common Homebrew locations, caches converted files outside the Messages attachment directory, keeps original_path stable, and reports converted_path / converted_mime_type separately for CAF→M4A and GIF→PNG conversions.

Proof: local make lint, make test (177 tests), make build, git diff --check; GitHub CI is green on 8dcb9d0 (run 25309469864). Closing this PR as superseded by main.