Poll the same watch receiver across monitor iterations instead of
cloning per pass; every clone re-observed the last write immediately
(~1.3M iterations/s measured over a relayed call), starving the
TURN/ICE read tasks.

Also: no_std ICE gather falls back to the get_local_ip seam for a
host candidate; add the dtls_monitor regression suite and the opt-in
`test-hooks` counter (DTLS_MONITOR_ITERATIONS).