Summary

• Cuts the steady-state cost of the PING/heartbeat loop: one batched IMAP
  STATUS round trip per poll iteration (instead of one per folder) and
  lock-free, memoized read-only state loads (zero SQL on unchanged
  synckeys).
• Makes SQL SyncCache saves dirty-field aware (merge instead of blob
  overwrite, skip when clean), matching the Mongo backend semantics.
• Batches email message fetches during export via a new optional
  Driver_Base::getMessagesBulk() (10 per batch, per-id fallback).

Implements the plan from #88. Companion PR: horde/Core#212.

Motivation

See #88: a device pinging a few dozen mail folders generated hundreds of
SQL statements and IMAP round trips per minute with zero changes to
report. Robustness first: every optimization is an optional hint with an
explicit per-item fallback, and foreign-client change detection is
unchanged (fresh STATUS data every iteration, prefetch consumed once).

Changes

• Collections::pollForChanges(): gathers the backend folder ids of all
  pinged email collections and calls the new no-op-by-default
  Driver_Base::prefetchFolderStatus() once per iteration; poll loads
  pass a new readonly flag through initCollectionState().
• Imap_Adapter::prefetchStatus(): single multi-mailbox STATUS
  (LIST-STATUS where the server supports it), consume-once in ping(),
  per-mailbox fallback on errors or missing/incomplete entries.
  getMessages() gains a uid_keys option.
• State_Base::loadState(): optional readonly flag — no collection
  lock, no garbage collection. State_Sql replaces SELECT … FOR UPDATE with a plain SELECT and memoizes decoded state rows per
  folder and synckey; any mutating load, state save, or synckey change
  invalidates the memo. Column metadata from Horde_Db is cached per
  table.
• State_Sql::saveSyncCache(): honors $dirty (per-collection merge,
  removals included), reads the stored blob in the query that previously
  only counted rows, UPDATE-first then INSERT, skips clean saves, and
  logs dirty property names instead of full cache blobs. SyncCache
  gets a missing dirty mark for bodypartprefs.
• Connector_Exporter_Sync: prefetches up to 10 upcoming CHANGE/DRAFT
  items through getMessagesBulk() (initial-sync bare-uid lists
  included); ids missing from a batch fall back to single getMessage()
  calls with unchanged error semantics (NotFound → drop and continue,
  TemporaryFailure → abort, other errors → keep batch in
  sync_pending).
• Design document doc/sync-performance.md (problem, invariants,
  fallbacks, observability, deferred work); cross-references added to
  architecture.md, integration.md, todo.md, and README.md.

Expected exceptions in new catch blocks: Horde_Imap_Client_Exception
(status prefetch), Horde_ActiveSync_Exception (folder-uid resolution in
the prefetch gather loop, bulk fetch in the exporter).

Test plan

• New unit tests: ImapAdapterTest (prefetch consumed by ping,
  consume-once, per-mailbox fallback on error),
  StateTest/Sql/ReadonlyLoadStateTest (lock-free load, memo hit
  without SQL, invalidation on mutating load and synckey change),
  StateTest/Sql/SyncCacheSaveTest (dirty merge preserves concurrent
  writers, per-collection removal, whole-property replace,
  skip-when-clean), Connector/ExporterPrefetchTest (bulk consumption,
  per-id fallback, no-bulk-support equivalence, initial-sync uid lists).
• Full package unit suite: no new failures (4 pre-existing failures
  unrelated to this change).
• Author deployment soak (iOS device): batched STATUS active,
  SYNC/PING behaviour unchanged, no errors in device logs.
• Review of the SQL query changes (@ralflang, as requested in RFC: Email sync performance — heartbeat loop state reloads, SyncCache writes, export batching #88).
• Multi-day soak with several concurrent devices (PING plus parallel
  SYNC on the same folders).