fix(ingest): run memorize in background when run_in_background=True

[relunctance]

Fix: ingest memorize async (#158)

Problem: POST /api/v1/ingest with run_in_background=True still blocks for 30-120s because only the memorize stage ran in background; add remained sync.

Root cause: The run_in_background flag was only passed to memorize(), not applied to add() which does the heavy LLM entity extraction + LanceDB write.

Solution: When run_in_background=True:

1. Run add() synchronously (memorize needs the dataset_id from add result)
2. Fire memorize() as asyncio.create_task() — return BACKGROUND_STARTED immediately
3. Log a warning if background memorize fails (fire-and-forget with error logging)

This unblocks hawk-memory's captureDualWrite() from timing out on mFlow calls.

Closes #158

[github-actions]

Welcome to the M-flow project, @relunctance! 🎉
Thank you for opening this pull request. A core maintainer will review your changes shortly. In the meantime, please make sure you've read our Contributing Guide.