[improve][broker] Use array-backed maps for Key Shared dispatch

[lhotari]

Motivation

Each Key_Shared dispatch batch creates two consumer-keyed maps: one groups entries selected for each consumer and the other tracks the consumer's remaining permits. These short-lived HashMaps allocate a node for every participating consumer and spend measurable CPU in HashMap.putVal while the dispatcher monitor is held.

Modifications

Use FastUtil Object2ObjectOpenHashMap for the two batch-local maps and size them from the connected consumer count. The maps keep the same Consumer.equals and hashCode semantics, values, and lifetime. Routing, permit accounting, ordering, and monitor scope are unchanged.

Add a focused JMH benchmark that constructs both maps and groups 1,000 entries across ten consumers.

Verifying this change

The focused benchmark measured:

Implementation	Time	Allocation
JDK HashMap	20.768 us/op	15,056.1 B/op
FastUtil open hash map	10.416 us/op	14,640.1 B/op
Change	49.9% faster	416.0 B/op less

A fanout screen used 500 producer connections, one topic, twenty Key_Shared subscriptions, and ten consumers per subscription. Two changed runs and three unchanged runs delivered every one of the expected twenty million fanout copies with zero duplicate, out-of-order, or invalid messages. Mean producer rates were 32,101 and 31,980 msg/s respectively, a flat 0.38% difference within run variation. The end-to-end runs are correctness and material-regression evidence; no whole-workload throughput gain is claimed.

Validation:

• PersistentStickyKeyDispatcherMultipleConsumersTest
• two-fork JMH run with GC allocation profiler
• ./gradlew quickCheck