feat(sabnzbd): paginate history via a single UNION query (no schema change)

[javi11]

What

Fix pagination of the SABnzbd-compatible mode=history API by moving the merge + dedup + pagination into one SQL query over the existing tables. No schema change, no migration.

Why

The handler merged three sources — completed rows from import_queue, a 7-day window of import_history, and failed rows from import_queue — deduped and sliced them in memory over hardcoded caps (2000 / 1000). So clients (Sonarr/Radarr) could never page past the caps, and noofslots didn't reflect the true total.

How

• New repo methods ListSABnzbdHistory + CountSABnzbdHistory: a UNION ALL of import_queue(completed) + import_history + import_queue(failed), deduped by an anti-join on import_history.nzb_id = import_queue.id (a completed import can live in both tables — the live queue row wins, the history copy is dropped), ORDER BY completed_at DESC, id DESC, real LIMIT/OFFSET, and an exact COUNT. Portable across SQLite + Postgres (no window functions, no basename/regexp; ? and FALSE handled by the dialect layer).
• Handler bulk path rewritten to use them; limit capped at 1000.
• SAB-history reports paths that do not exist on disk → ARR queues fill with FileNotFoundException #596 preserved — only rows sourced from the live completed queue are flipped to Failed when their reported path is missing on disk.
• nzo_ids reconciliation path unchanged.

What does NOT change

• No new tables/columns, no migration.
• Failures keep living in import_queue with their existing FailedItemRetentionHours lifecycle (default 24h) — same as before.
• Completed/failed queue rows, retention config, the UI, and Stremio are all untouched.

Behavior note

Dedup shifts from filename-based to queue-id-based (import_history.nzb_id). This is more correct than the old name dedup (which was partly ineffective because queue paths carry a {id}- prefix that import_history.nzb_name strips): a completed import in both tables now reliably shows once, and two genuinely distinct jobs that share a filename both show (as real SABnzbd does). The old fixed history window (7 days) is gone — pagination now spans the full history.

Test plan

• go test ./internal/database/... ./internal/api/... — green; go vet clean; go build ./... clean
• New tests: dedup across queue+history (single row, correct source), failed-item inclusion, skip_arr exclusion, category filter, paged enumeration with no overlap/gaps and newest-first ordering, start beyond total
• Manual: mode=history with start/limit (no page overlap, constant noofslots across pages)