perf(ingest): reduce database work and buffer MQTT bursts

[MrAlders0n]

Changes

• Move packet processing out of the MQTT callback into bounded workers, keeping messages from each observer ordered.
• Cache successful capability writes and skip endpoint lookups for suppressed copies.
• Log queue overflow, release rejected MQTT messages, and drain workers before the final presence flush.

Validation

• Build, formatting, vet, full tests with PostgreSQL, and ingest/presence/hub race checks passed.
• Local load tests delivered all 15,000 live hearings and stored all 5,000 observations at 2,000 MQTT messages/sec.
• With added database latency, the baseline delivered 460/3,000 hearings; this candidate delivered all 3,000 without drops or reconnects.
• Initial changes reduced Beacon CPU by 21% in the original synthetic workload. In a separate mixed workload, skipping unused endpoint lookups reduced CPU another 6% against the prior candidate (three-run medians).

Buffers remain bounded. Sustained overload can still drop messages; this does not add durable delivery. No API or schema changes.