Redirect/pool: endpoint matching has no normalization, newest hint not preferred, cache TTL/entries not configurable, TableSessionPool lacks acquire_for_device

Smaller redirect/pool issues around `acquire_for_device`:

1. **Redirect endpoint matching is exact host-string equality with no normalization**. Hostname-vs-IP spelling silently disables the optimization on multi-DataNode clusters, with no log line so it is invisible from outside. At minimum the comparison should normalize case/whitespace and resolve hostname-vs-IP where feasible.
2. **When idle sessions hold conflicting hints for one device, the winner is queue position, not recency.** `RedirectCache` already tracks `seq`/insertion order — prefer the newest hint (this stays within the documented "any idle session" behaviour).
3. **`RedirectCache` TTL and `max_entries` are fixed at construction** with no `SessionConfig` knob; this becomes measurable on large batch inserts. Add `SessionConfig` fields with the current defaults.
4. **`TableSessionPool` exposes `acquire()` but not `acquire_for_device()`**, so table-mode users accumulate hints no pool API can consume. Expose the underlying `SessionPool::acquire_for_device`.

Start by locating RedirectCache, SessionConfig, TableSessionPool, and SessionPool::acquire_for_device, then trace how acquire_for_device selects idle sessions and how redirect hints are compared. Check the existing pool and redirect tests, and verify completion against all four requested behaviors: normalized matching, newest-hint selection, configurable current defaults, and the table-pool API.

Written by the indexing model from the issue text.