[matthewbayer]

Reduce shared-registry contention while preserving per-thread buffer reuse.

DefaultByteBufferPool acquires a shared map monitor to find a thread’s cached buffers on allocation and release. Threads contend on that monitor even when reusing their own buffers.

This change gives each allocating thread a weak reference to its existing cache:

• Most cached lookups use a Java ThreadLocal. The synchronized weak-key registry continues to own the cache data and support cleanup from another thread.
• Every 256 successful cached lookups, the registry is accessed to expunge dead threads’ entries. Weak references prevent long-lived workers from retaining discarded pools.
• Pools with caching disabled bypass the added ThreadLocal lookup and retain their existing registry/queue path.

Jira: https://redhat.atlassian.net/browse/UNDERTOW-2809

[matthewbayer]

@fl4via curious for your thoughts here! We experience pretty significant degradation under high concurrent request counts due to this lock, happy to help out here in any way to address this problem.