…nitialize() epoch-gate

Fixes a rare CI fatal AccessViolationException at LogRecord.get_Info /
ObjectScanIterator.GetPhysicalAddressAndAllocatedSize during
VectorManager.RunCleanupTaskAsync. Same crash signature as PR #1765
but a different remaining gap.

## Root cause

PR #1765 made AllocatorBase.Reset()'s page-free section epoch-safe
(2-phase BumpCurrentEpoch barrier around OnPagesClosed +
FreeAllAllocatedPages). It did NOT cover the per-allocator Initialize()
that runs after base.Reset() returns. Initialize() non-monotonically
rewinds HeadAddress / BeginAddress / TailPageOffset back to
FirstValidAddress while pagePointers[i] are mostly 0 (only pages 0 and 1
are reallocated). Race window between HeadAddress reset and
TailPageOffset reset:

  - Iterator reads stopAddress = TailAddress = old high (stale)
  - Iterator passes currentAddress < stopAddress check, proceeds
  - Iterator reads headAddress = HeadAddress = new low
  - Iterator chooses in-memory path (currentAddress >= headAddress)
  - Iterator dereferences pagePointers[X] + offset = 0 + offset → AVE

The cluster re-attach paths (ReplicaDisklessSync, ReplicaDiskbasedSync)
call SuspendPrimaryOnlyTasksAsync() AFTER Reset, and even then it only
covers TaskManager-registered tasks. VectorManager.cleanupTask is
started directly in the constructor (VectorManager.cs:110) as a plain
async Task and is invisible to TaskManager.

## Fix

Two complementary fixes:

### A. Tsavorite-side defense-in-depth

- AllocatorBase: added `internal volatile bool Initializing`.
  Refactored Reset() to be non-virtual, wrapping the existing 2-phase
  epoch logic (renamed to ResetCore()) PLUS the per-allocator
  Initialize() call inside Initializing = true / false (try/finally so
  a throw can't leave iterators stuck).
- Per-allocator Reset() overrides removed in ObjectAllocatorImpl,
  SpanByteAllocatorImpl, TsavoriteLogAllocatorImpl — base now does
  ResetCore() + Initialize() directly under the gate.
- ObjectScanIterator + SpanByteScanIterator (InitializeGetNextAndAcquireEpoch):
  - Move epoch.Resume() BEFORE sampling TailAddress (was sampled outside
    epoch protection — preexisting issue).
  - Spin-wait `while (hlogBase.Initializing) { Thread.Yield(); epoch.ProtectAndDrain(); }`
    with `!assumeInMemory` bypass (avoids deadlock with internal
    MemoryPageScan inside OnPagesClosed which constructs assumeInMemory
    iterators on the same thread that is running the Initializing-gated
    deferred action).
- LoadPageIfNeeded: recheck `currentAddress >= stopAddress` after
  snapping forward to BeginAddress (stale snap could place
  currentAddress at TailAddress with an unparseable record).

### B. Caller-side explicit drain (restores Reset's documented contract)

- VectorManager.Cleanup.cs: added PauseCleanupAsync() / ResumeCleanup()
  API. Cleanup loop uses increment-then-check pattern (bump
  cleanupInProgress first, re-read pauseCleanupRequests, back out if
  non-zero) so a pause caller cannot race past an iteration about to
  start.
- ReplicaDisklessSync.cs + ReplicaDiskbasedSync.cs: wrap
  storeWrapper.Reset() in `await vectorManager.PauseCleanupAsync()` /
  `vectorManager.ResumeCleanup()` try/finally.

A is defense-in-depth (any future background reader is safe); B
restores the documented contract for the known case and skips the
per-iteration spin-wait when callers cooperate.

## Validation

- Build: 0 warnings, 0 errors. dotnet format clean.
- Tsavorite tests: 1453/1453 pass.
- Garnet.test Reset/Recovery/Iterate filter: 30/30 pass.
- ClusterMigrate + ClusterVectorSet (52 tests): all pass.
- MigrateVectorSetWhileModifyingAsync (the originally failing cluster
  test): stable across 8 sequential runs (~18s each) with --blame-crash.
- VectorCleanupVsResetRaceTests (PR #1765's regression): pass [Repeat(5)].
- ClusterResetDuringReplicationTests: 2/2 pass.

Co-authored-by: Copilot <[email protected]>