Fix channel backpressure causing keepalive timeout and stream drops

[joshkautz]

Problem

When the internal mpsc message channels reach capacity, send().await blocks the entire message processing loop in Poller::run(). While blocked:

1. Keepalive ping messages from the camera cannot be processed
2. The camera interprets the silence as a dead connection
3. The camera drops the session after its ping timeout expires
4. Neolink enters a reconnection cycle

This is the root cause of streams dying after hours of continuous operation, which users work around with periodic cron restarts.

Additionally, when RTSP clients cannot consume frames fast enough, audio buffer overflow triggers FlowError::Flushing, which cascades into the video pipeline causing a full stream reconnection.

Fix

1. Increase channel capacities to reduce how often channels fill under normal load:

• Outgoing message channel: 100 → 500
• Poll command channel: 200 → 1000
• Subscriber channels: 100 → 500

2. Non-blocking send when channel is full. When sender.capacity() == 0, use try_send() instead of the blocking send().await. Dropping a single video frame is far better than blocking the message loop and losing the entire camera connection. The normal async send path is preserved when capacity is available.

3. Drop audio frames on buffer overflow. When the audio AppSrc buffer exceeds 90% capacity, audio frames are silently dropped instead of being pushed. This prevents cascading backpressure from stalling the video pipeline. Audio buffer overflow was triggering pipeline flushing, which caused full stream reconnection cycles.

Changes

• crates/core/src/bc_protocol/connection/bcconn.rs: Increase channel sizes, replace blocking send with try_send at capacity
• src/rtsp/factory.rs: Add audio frame dropping when AppSrc buffer is near capacity

Relationship to Other PRs

The audio frame dropping in this PR is complementary to #394, which adds a config option to disable audio entirely. The two approaches serve different use cases:

• feat: audio and long latency as config #394: Hard disable — audio pipeline is never constructed. Best for users who know they don't need audio and want to save CPU permanently.
• This PR: Graceful degradation — audio passes through normally but is dropped when the buffer is under pressure. Prevents cascading failures without requiring any configuration change.

Both could be merged. Users who disable audio via #394 would never hit the audio drop path in this PR.

Related Issues

• Endless "buffer full" #349 — Endless "buffer full" spam making streams unusable
• Consistently Failing to send to Frigate. V0.6.2 #346 — Consistent stream drops requiring periodic restarts
• Neolink crashing #315 — "Failed to send to source" and "Buffer full on audsrc" errors
• Neolink V0.6.3 - Memory Leak #366 — "Reaching limit of channel: Remaining: 0 of 100" warnings

[joshkautz]

Note: These are tweaks I ended up making for my own use case — a 24/7 ALPR system where streams need to stay up indefinitely and audio isn't critical. If the nature of these changes doesn't align with the trajectory or direction you want to take the project, feel free to close this — I'm simply opening it because they solved real stability issues for me and might help others in similar situations.