Improve replay throughput and asynchronous event fan-out

[jillesvangurp]

Summary

This PR improves end-to-end replay throughput by taking more work off the synchronous ingest path and by reducing downstream event delivery overhead.

What Changed

• Introduces an explicit two-stage async ingest pipeline in the hub:
  • a native publication stage for native location and motion fan-out
  • a decision stage for transform-dependent follow-on work such as alternate-CRS publication, geofence evaluation, and collision evaluation
• Keeps transforms lazy so they are only computed if a later stage actually needs them.
• Adds a stable insertion point for future filtered track processing before fence/collision decisions without implementing filtering yet.
• Adds runtime drop counters for overloaded non-critical paths.
• Raises and externalizes queue sizing for event-bus subscribers, native publication, derived/decision processing, and WebSocket outbound buffering.
• Changes WebSocket fan-out to batch internal events before encoding wrapper payloads.
• Changes WebSocket slow-consumer behavior from immediate disconnect to dropping outbound payloads on overflow.
• Adds a cheap collision prefilter based on active motions and bounding checks before full collision evaluation.
• Batches replay connector WebSocket publishes with configurable batch window and maximum batch size.
• Updates docs and replay connector configuration to describe the new pipeline and batching controls.

New Or Updated Knobs

• EVENT_BUS_SUBSCRIBER_BUFFER
• NATIVE_LOCATION_BUFFER
• DERIVED_LOCATION_BUFFER
• WEBSOCKET_OUTBOUND_BUFFER
• REPLAY_BATCH_WINDOW_MS
• REPLAY_MAX_BATCH_SIZE

Benchmark Result

Using the same captured OpenSky trace replayed with --interpolation-rate-hz 10.0:

• scheduled emissions: 13,192
• previous observed replayed location items: 8,639
• new observed replayed location items: 13,192
• fence events remained correct at 126, matching the live capture
• WebSocket wrapper messages dropped from 8,640 to 1,409 because of batching

Validation

• just bootstrap
• just generate
• just check
• rebuilt the local demo stack and reran the OpenSky replay benchmark against the updated hub