ZOOKEEPER-2466: Fix C client skipping servers when reconnecting

[kosa-as]

Problem

The C client can advance the server address cursor twice for a single
failed connection attempt.

After connection failures, handle_error() advanced the address cursor,
while the next reconnect attempt also selected the next server through
zoo_cycle_next_server(). Once the client falls back from reconfig server
selection to the normal round-robin list, this can skip servers and may
cause the client to keep retrying only part of the server list.

Fix

Keep normal reconnect server cycling in zoo_cycle_next_server() and
remove address cursor advancement from handle_error().

Move the reconnect delay calculation next to the server cycling logic so
the delay is still set when the normal server list wraps.

Also guard read-only mode's read/write server probing cursor advancement
with the existing reconfig lock, so it does not advance the normal address
list while reconfig server selection is active.

Testing

Added a regression test for ZOOKEEPER-2466 by exhausting the configured
server list while all servers are unavailable, then starting one server
and verifying the same client eventually reconnects to it.

Verified locally with a behavior harness:

• unpatched client: does not reconnect after the server starts
• patched client: reconnects successfully