[RangeIndex] Replace busy-spin snapshot claim with SemaphoreSlim to avoid CPU spikes

[tiagonapoli]

Problem

The per-tree snapshot claim in TreeEntry serialized concurrent CPR snapshots — OnFlush (flush), SnapshotAllTreesForCheckpoint (checkpoint), and SnapshotForMigration (cluster migration) — using a CAS + busy-spin:

while (!TryClaimSnapshot())   // Interlocked.CompareExchange on an int
    Thread.Yield();

However, in case of pretty big BFTree, and since we may be writing multi-gigabyte files, snapshot may take more time, so any waiter will burn CPU for the entire wait. If migration happens during a checkpoint, or flushes happen with migration/checkpoint, we might burn CPU unnecessarily.

Fix

Replace the CAS + spin with a per-tree SemaphoreSlim so waiters block instead of spinning:

snapshotLock.Wait();
try { /* injected latency hook */ CprSnapshotByPtr(...); }
finally { snapshotLock.Release(); }