Client seed addresses: live failover and routing cost

The change touches only the connect, reconnect and slot map reload paths of the six clients (Client behavior); the server is unchanged.

scripts/seed_failover.py starts two standalone servers holding seed:name = a and b, runs a client probe with both as seeds, stops the first server and checks that the probe reads b afterwards. Debug server build, macOS (Apple silicon), 2026-10-19.

Client	Probe	Result
Rust	clients/ruvio-client/examples/seed_probe.rs	first a, after b
C#	clients/csharp/Ruvio.Client.SeedProbe	first a, after b
Go	clients/go/internal/seedprobe	first a, after b
Java	clients/java/src/test/java/.../SeedProbe.java	first a, after b
Node	clients/node/bench/seed-probe.ts	first a, after b
Python	clients/python/benchmarks/seed_probe.py	first a, after b

Before this change no client took a second address, so the same stop left every client retrying the stopped server.

Each client’s routing-cost tool (Cluster routing cost in the clients: before and after) against one release-fast server with 4 shards and port_mode = "per-shard", FLUSHALL before each run, the base commit and the change one after the other, one run each. User CPU nanoseconds per request, base → change:

Client	one-node GET P=16	one-node SET P=64	cluster GET P=64	plain GET P=16	allocations / bytes per request
Rust	145 → 146	141 → 144	215 → 269	109 → 117	same
C#	705 → 798	260 → 255	699 → 645	1026 → 833	same or lower
Go	483 → 320	157 → 155	316 → 339	482 → 446	same
Java	324 → 332	148 → 120	340 → 342	246 → 254	same or lower
Node	586 → 728	336 → 390	638 → 358	621 → 785	not counted by the tool
Python	1384 → 1354	1174 → 1169	1088 → 1070	758 → 740	not counted by the tool

Other benchmarks were running on the same machine during the Rust, C#, Go, Java and Node runs (load average about 10), so their CPU numbers move in both directions, including rows whose code did not change (plain). The Python run was taken alone and moves by 2% or less. Allocations and bytes per request did not grow in any client. A quiet repeat on a dedicated machine was not made.