Reduce notification timeout long tail

[raymondjacobson]

Summary

• use a smaller candidate window for timestamp pagination so old notification scrollback does not read tens of thousands of historical rows before grouping
• give unread polling its own shorter timeout and return a conservative capped unread response instead of surfacing a 504 when the multi-recipient array path is cold

Production evidence

• slow dEKBy cursor with the existing 25k candidate window read ~23.4k rows and built ~2.9k groups before trimming
• the same cursor with a 2k candidate window still produced >600 groups and planned/executed in ~51ms warm during read-only production probing
• unread timeout case for user 843 was dominated by the multi-recipient GIN path scanning ~42k historical rows to find 7 current unread rows, so this PR fails closed for the poll while we plan the recipient-normalized/index follow-up

Validation

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

Notes

• go test ./api -run TestV1Notifications is blocked locally because Postgres on localhost:21300 refused connections