Various Cuda/FFI Memory Lifecycle Fixes

[alan-george-lk]

Pulls in livekit-ffi/v0.12.77, fixing the following issues:

• Room disconnect closes peer connections before signal teardown, so cancelled close can no longer leak ICE UDP sockets
• Repeated connect/disconnect no longer retains room sessions or data-channel observer cycles
• Dropping an unused VideoSource now releases its keepalive buffers instead of pinning them for the process lifetime
  NVIDIA encoder/decoder factories tear down a shared CUDA context on last drop, and NVENC init failures no longer leave a half-initialized encoder
• FFI shutdown drops leftover handles one at a time so nested dispose cannot re-enter handle-map clear
• PlatformAudio runs all ADM access on the WebRTC worker thread, creates the ADM lazily, and keeps Android recording and shutdown races from dropping audio
• Linux dmabuf frames are captured through the existing video capture path
• Publishing pre-encoded video on macOS no longer segfaults

[devin-ai-integration]

Devin Review found 1 potential issue.

🐛 1 issue in files not directly in the diff

🐛 Concurrent shutdown leaks FFI resources

A request storing a handle after dispose snapshots the map escapes cleanup. store_handle does not synchronize with disposal. The resource survives until another shutdown or process exit.

1 flag not posted on this PR by your GitHub settings — view it in Devin Review. (Configure)