Make Vector Set recovery re-entrant so disk-based full sync sanitizes handles

[tiagonapoli]

Stacked on #1 — review that one first; this PR targets tiagonapoli/vs-1-diskless, so its diff shows only this change.

Problem

VectorManager collects the contexts it sees during recovery into recoveredIndexes / recoveredMetadata, and ResumePostRecovery reconciles them against the recovered store. Both were released at the end of that method, on the assumption that a process recovers exactly once.

A replica does not. It re-recovers the store on every disk-based full sync, and nothing re-ran ResumePostRecovery on that path.

So the second recovery threw NullReferenceException out of RecoveredVectorSetIndexKey. That exception escaped ClearBitsOnPage — the per-record walk that fires OnDiskRead — so the walk aborted after the first record on the page. Every Vector Set record behind it kept the primary's raw DiskANN handle.

Instrumented recovery of three vector sets, before the fix:

CBOP page=0 startLA=0 endLA=16777216 until=156296 endOff=156296 snapFrom=64
  rec off=64   size=8224  ->  msf_            (context metadata)
  rec off=8288 size=88    ->  {vsdisk}dblarge (sanitized OK)
  EX System.NullReferenceException
     at VectorManager.RecoveredVectorSetIndexKey(LogRecord&) VectorManager.cs:384
     at GarnetRecordTriggers.OnRecoverySnapshotRead(LogRecord&)
     at TsavoriteKV.ClearBitsOnPage(...) Recovery.cs:1259

dbsmall and dbmedium were never visited, and came up on the replica holding IndexPtr == primary's pointer.

The half-finished recovery also leaked the checkpoint file handle, which is what made teardown hang ~60s deleting the directory. The disk-based fixture went 1m07s → 7s.

This path matters more than the diskless one in practice, since ReplicaDisklessSync defaults to false.

Fix

• Recovery state is cleared rather than released, so a second recovery starts from a clean slate instead of a null reference.
• A replica's disk-based recovery reconciles it via the new StoreWrapper.RecoverVectorSets(), mirroring what the startup path already did.

Validation

• 16/16 pass with the fix.
• All 4 disk-based tests fail without it.
• Regressions green: Garnet.test.cluster.vectorsets 80/80, replication.disklesssync 28/28, replication 103/103.