SessionPool acquire: growth/hand-out failures bypass acquire_timeout budget; live counter decremented outside the state lock loses Condvar wakeups; acquire_timeout = Duration::MAX panics

Three related pool-accounting defects found by stress-testing `SessionPool` against fake listeners:

1. **`acquire()` spends none of its `acquire_timeout` budget when the growth branch fails.** When `open_session()` fails, `acquire()` returns the error immediately (measured 198us against a 5s budget; 4136 of 4800 acquires failed instantly under stress while sessions were circulating). The same happens when the `USE` replay in `hand_out` fails — it discards the remaining idle candidates instead of retrying the loop.
2. **`live` is decremented outside the `state` lock at several sites**, and the decremented value is exactly the predicate parked waiters re-evaluate, so Condvar notifications can be lost. The comment "Only mutated while holding `state`" no longer matches the code. Measured stalls track `acquire_timeout` exactly (300ms → 305ms, 1000ms → 1.005s).
3. **`acquire_timeout = Duration::MAX` panics** on `Instant::now() + acquire_timeout` overflow before the lock is taken. There is no "never time out" option, and `Duration::MAX` is the natural way to ask for one.

Fix: on growth/hand-out failure, re-take the lock, decrement `live` under it, notify, and continue the loop (bounded by the original deadline); use a saturating deadline (`checked_add`) so `Duration::MAX` means "wait without timeout".

Start by locating SessionPool::acquire and tracing the open_session growth path and hand_out USE replay path, then inspect every live mutation relative to the state lock and Condvar notification. Verify that failures retry until the original deadline, live decrements notify waiters under the lock, and Duration::MAX waits without overflow; reproduce the behavior with fake-listener stress tests.

Written by the indexing model from the issue text.