feat: coalesce repeated log messages to prevent log floods

[tobez]

Summary

Repeated per-agent SNMP anomalies and per-connection client churn could
previously drive log output without bound — a misbehaving or hostile agent
could flood the log with thousands of lines per second. This adds windowed
coalescing: the first event per key in a 10-second window logs immediately
with full detail; the rest are counted and rolled up into a single
… repeated=N interval_s=10 line when the window closes.

Storm cost drops to at most one detailed line plus one rollup per 10 s per
key. The UDP send-buffer overflow, previously dropped silently, now gets a
throttled notice so operators can tell sends are being lost; its statistics
counter stays exact and unconditional.

Details

• Pure clock-injected coalescing primitive, unit-tested with a fake clock.
• One category per distinct message; per-destination counters are lazily
  allocated only for hosts that actually misbehave, so memory scales with
  misbehaving hosts, not fleet size.
• Rollups piggyback the existing once-per-second housekeeping tick — no new
  timer machinery. log_debug output is left unthrottled (it's opt-in).

Test plan

• New C unit tier t/throttle.t (fake-clock primitive + glue), 21 assertions
• t/logging.t extended: per-destination anomaly coalescing + client-churn coalescing subtests
• Full suite: 343 tests pass (make test)
• scan-build clean — No bugs found
• Clean -Werror build