Reclaim idle connections across routes in RouteSegmentedConnPool - #722

RouteSegmentedConnPool stopped allocating when the global limit was reached, even if another route held an idle reusable entry, and its hot path paid for two shared structures: the executor hand-off of every graceful close and the single deque of idle entries of each route.

Three commits:

1. Reclaim idle entries across routes at max capacity. When maxTotal is reached, an idle entry of another route is reclaimed and its global slot transferred to the requesting route. The hot path stays route-local; pending routes are retried when capacity appears.
2. Lock-free disposal and drain hand-offs. Graceful closes go through a bounded lock-free queue served by on-demand daemon threads: a caller wakes a worker only when none is awake, idle workers linger in short timed parks, and workers bring in more threads themselves when their closes are slow or the backlog grows. close() never waits on the backlog. The round-robin drain runs on a single on-demand thread signalled with a CAS. Segments are retired with a negative allocation marker before leaving the map, which fixes a global slot leak when an allocation raced the removal of a segment.
3. Per-thread idle entries. Under LIFO the idle entries of a route are kept in independent Treiber stacks, one per stripe, with the heads on separate cache lines; a thread pushes to and pops from its own stripe and steals from another only when it is empty. Reuse is most recent per thread rather than per route, and eviction takes the least recently released entry of one stripe. FIFO keeps the deque.

Benchmarks

JMH 1.37, JDK 17, 16 cores, -f 2 -wi 3 -w 1 -i 5 -r 1, each suite run alone. RouteSegmentedConnPoolBenchmark has no I/O and 8 threads unless stated; RoutePoolsJmh drives real sockets against a local HTTP/1.1 server with 50 threads, maxTotal=100, 5 per route, and reports completed requests.

Before these changes OFFLOCK measured 3.6 M ops/s on hotRouteUncontended, 1.9 M on coldRouteAtFullPoolWithIdle and 1.4 µs on the backlog case, and the socket benchmark with slow closes did not complete. hotRoute with a single connection and the instant-close socket rows are bound by thread hand-off latency and by the server respectively, so all three pools measure the same there. STRICT closes connections inline, which is why it leads the backlog latency with no-op closes and collapses with real ones.

httpcore5: 2007 tests, Checkstyle clean.