[client] Fix unbounded growth of stale partition entries in writer metadata cache

[lilei1128]

Purpose

Linked issue: close #xxx

When partitions are continuously created and dropped, physical table path
entries in RecordAccumulator.writeBatches and Cluster.partitionsIdByPath
were never removed, causing unbounded memory growth and CPU waste from
iterating over stale entries in ready() and drain() hot loops.

Brief change log

Fix 1 (RecordAccumulator): introduce markPathsAsStale() and
removeStalePathIfEmpty() with a dedicated staleLock so that the
check-then-remove is atomic with respect to concurrent append() calls.
New appends to stale paths are rejected immediately.

Fix 2 (MetadataUtils): during partial metadata updates, remove stale
partition entries for any table whose partition list was refreshed, so
that dropped partitions no longer linger in partitionsIdByPath.

Fix 3 (Sender): call cleanupStaleWriteBatches() on every sendWriteData()
cycle to mark paths absent from the cluster as stale and remove them once
their deques are fully drained.

Tests

• RecordAccumulatorTest.testMarkAndRemoveStalePathAfterDrain: verifies
  that append() is rejected after markPathsAsStale(), and that
  removeStalePathIfEmpty() succeeds once the deque is drained.
• RecordAccumulatorTest.testRemoveStalePathReturnsFalseWhenDequeNonEmpty:
  verifies that removeStalePathIfEmpty() returns false and preserves the
  path when pending batches remain.
• SenderTest.testSenderCleansUpStaleWriteBatchesAfterMetadataUpdate:
  verifies end-to-end that runOnce() removes a stale path from
  writeBatches after the cluster no longer contains it and its deque
  has been drained.

API and Format

Documentation

[lilei1128]

close #3362