Send audio to the inference process in ~40ms blocks instead of 0.5s, so
end-to-end latency is bounded by the streaming model (~180ms) rather than IPC
batching. Implement the chunked downmixing path in DataCallback (slice-wise,
RT-safe) instead of crashing, and stop clearing the graph-thread scratch
buffer from the main thread in Stop() (a data race the lower latency exposed).
Count the amount of frame dropped, increase ringbuffer depth to smooth
over load spike events.