[bardesss]

Read-state fixes in the offline mark queue and sync. Stacked on #36; merge that first.

Queue

• A mark-all now takes over the pending read marks it covers (same feeds, published at or before its cutoff). Marking one of those items unread afterwards queues a real unread mark. It no longer cancels the read mark and leaves the server with the item read.
• PendingMarkEntity gets an attempted flag (DB version 4, auto-migration, 4.json committed). The flush sets it before sending. A reversal cancels only a row that was never attempted, so a send that timed out after reaching the server is corrected instead of lost. The flag survives when a row is replaced.
• A mark-all the server rejects with a 4xx (other than 401, 403, 408 and 429) is dropped, so it no longer blocks every item mark behind it. 5xx, 408 and 429 keep it queued.

API and sync

• The edit token is fetched again on 400 and 403, as well as on 401, once per call.
• Paging stops when a page adds no new id or repeats its continuation. That fetch counts as incomplete.
• A flush-only job returns retry() with exponential backoff from 30 s, up to 5 attempts, while marks stay queued.
• unreadIds for folders with more than 500 feeds reads all unread rows and filters them, keeping the order, instead of going over SQLite's argument limit.

One commit adds a publish-time window for marking items read on a capped refresh, and a later commit reverts it. Servers sort the unread listing by fetch time, not publish time, so that window could mark items read that are still unread.

Tests: MarkQueueTest, MigrationTest (3 to 4), GReaderApiTest, SyncEngineTest, SyncWorkerTest, ItemDaoTest. Full unit tests and lint pass.