incr/decr operations bypassed the ensureWriteQueueSize check that other
  write paths (set, add, touch) use, causing them to block on unreachable
  or stalled nodes for the full mutateOperationTimeout duration. A single
  unhealthy node could cause all incr/decr calls to block, exhausting the
  caller's thread pool.

  This adds the same ensureWriteQueueSize gate to incr/decr. If the target
  node is inactive or its write queue is full, the operation returns -1
  immediately, consistent with the documented API contract. The existing
  reconciliation logic in EVCacheImpl.incr()/decr() already handles -1 as
  a zone failure and will sync the value from healthy zones on the next
  successful operation.

… write path

- EVCacheImpl.decr: fix reconciliation to pick the minimum non-(-1) value
  across nodes instead of the maximum. For decr, the most up-to-date node
  has the lowest value (most decremented), so the old max-pick logic would
  overwrite correctly decremented nodes with stale higher values.

- EVCacheClient.ensureWriteQueueSize: add isAvailable() check before
  entering the retry/sleep loop. This fast-fails writes to inactive nodes
  instead of blocking request threads through 3 retry iterations. Affects
  all write operations that go through ensureWriteQueueSize (incr, decr,
  set, delete, etc.).