Refactor: split process_attachments.py into focused modules #74

Context

`src/maildb/ingest/process_attachments.py` is ~750 lines and houses five distinct concerns:

1. The extraction claim/process loop
2. Embed-chunk pipeline
3. Empty-extraction sweep + reembed
4. Subprocess worker entrypoint
5. The supervisor

Grew organically across PRs #54, #56, #58, #60. Now it's hard to hold the whole thing in head while reading any single function.

Proposal

Split into:

• `process_attachments/init.py` — re-exports for back-compat
• `process_attachments/claim.py` — `_claim_row`, `_load_attachment`, `_set_status`, `ensure_pending_rows`, `_reclaim_stale`, `_count_selected`
• `process_attachments/embed.py` — `_embed_chunks`, `EmbedFailedError`, `_build_embedding_client`, `reembed_zero_vectors`, `sweep_empty_extractions`
• `process_attachments/pipeline.py` — `process_one`, `_claim_and_process_loop`
• `process_attachments/supervisor.py` — `_run_supervised_single_worker`, `_find_stuck_extracting`, `_killpg_quietly`, `_mp_context`
• `process_attachments/run.py` — `run`, dispatch logic, `_subprocess_worker`

Each module < 200 lines, each tested in isolation. No behavior change.

Acceptance criteria

• Module split per the layout above.
• All existing tests pass without modification (back-compat exports).
• No public API change to callers (CLI, MCP).

🤖 Generated with Claude Code