Wake blocked checkout callers when a connection is discarded

[OskarEichler]

Summary

Wake blocked checkout callers when discarding a connection releases capacity. TimedStack#decrement_created now updates the count under the existing mutex and broadcasts to waiting callers, like the normal checkin path.

Previously, a size-one pool could discard its only connection while another caller remained asleep until the full checkout timeout, despite replacement capacity being available.

Reproduction

require 'connection_pool'
pool = ConnectionPool.new(size: 1, timeout: 0.5) { Object.new }
ready = Queue.new
worker = nil
pool.with do
  worker = Thread.new do
    ready << true
    started = Process.clock_gettime(Process::CLOCK_MONOTONIC)
    pool.with { puts Process.clock_gettime(Process::CLOCK_MONOTONIC) - started }
  end
  ready.pop
  Thread.pass until worker.status == 'sleep'
  pool.discard_current_connection
end
worker.join
pool.shutdown { |_connection| }

On installed 3.0.2: ~0.502 seconds. With the fix: under 0.001 seconds in 20 local rounds. These are an isolated contention reproduction, not a general throughput benchmark.

Verification

• Ruby 4.0.6: existing suite passes, 86 runs / 187 assertions / zero failures or errors on the release-based patch.
• Focused repeated wakeup and pool-capacity checks pass.
• Existing targeted StandardRB, Ruby syntax and git whitespace checks pass.
• Consumer patch based on v3.0.2; this PR carries only the corresponding focused change on main.

Limitations and breaking changes

• The release's development setup resolves Minitest 6 without its extracted mock support. Initial suite had four missing-stub errors. An external verification Gemfile adds minitest-mock, as current upstream does. No test, Gemfile or gemspec changes included.
• No new tests under the consumer's no-new-tests policy; the reproduction and existing suite provide verification.
• No JRuby/TruffleRuby/Linux runtime checks were run locally; CI should exercise the existing matrix.
• Breaking changes: none intended. Public API, dependencies and ordinary checkin behavior are unchanged. Waiters can now use replacement capacity promptly.

Prepared with Codex assistance.