Optimize notification unread polling

[raymondjacobson]

Summary

• Gate limit=0 notification unread polling behind the fast single-recipient path before touching the broader multi-recipient GIN scan.
• Add a concurrent partial GIN index for multi-recipient notification.user_ids rows.
• Cover capped unread count and single/multi group dedupe behavior in tests.

Live analysis

• Current bridge reads; replication lag was ms-level during the investigation.
• Remaining /v1/notifications/:userId 504s were low-volume but concentrated in a few users, with requests timing out at the 8s notification read deadline.
• Read-only EXPLAIN against the replica for top user DvEvv/77214 improved the unread poll shape from roughly 132ms / 39,651 buffer hits to roughly 14ms / 1,132 buffer hits by avoiding the multi-recipient branch when the single-recipient branch already satisfied the capped unread count.
• User eJ57D/51 showed the same branch-elision behavior.
• Residual users with fewer than 20 single-recipient unread groups still need the fallback, so the partial multi-recipient GIN index targets that path.

Testing

• go test ./api -run '^$'
• git diff --check

Note: focused notification integration tests could not run locally because the local test Postgres at localhost:21300 was not running.