feat(invoke): reuse existing InvokeAI images instead of re-uploading

[lstein]

Summary

• When the selected reference image looks InvokeAI-managed (UUID-style filename, or raw metadata with app_version / generation_mode / canvas_v2_metadata), probe GET /api/v1/images/i/{name} and skip the upload when InvokeAI already has the file — the recall is sent with the existing filename instead.
• Non-Invoke images (no UUID name, no Invoke metadata) keep the original upload-then-recall path, so there's no extra round-trip for ordinary photos.
• Probe failures (connection errors, auth issues, anything non-200) fall through to upload, so the optimization never breaks a request that would otherwise succeed.
• Response payload now includes reused_existing: bool for observability; frontend already only checks response.ok.

Test plan

• pytest tests — 134 backend tests pass, including 5 new use_ref_image tests covering UUID hit, UUID miss (404 → upload), metadata-based probe, non-Invoke skip, and probe-error fallthrough
• npm test — 239 frontend tests pass
• make lint — ruff + eslint + prettier clean
• Manual: pick an image that originated from InvokeAI (UUID filename), click "Use as Ref Image", confirm recall succeeds with no duplicate appearing in the target board
• Manual: pick a non-Invoke image (e.g. a phone photo), confirm upload-then-recall still works
• Manual: with InvokeAI briefly unreachable mid-flow, confirm the request still falls through to upload cleanly

🤖 Generated with Claude Code