fix(pulseaudio): release server-side stream on Stream::drop

[knz]

The PulseAudio backend spawns a driver thread that calls pulseaudio::PlaybackStream::play_all, which awaits source EOF before draining and exiting. EOF for a PlaybackSource is signalled by poll_read returning 0, but the data-callback wrapper used by new_playback always reports the full buffer length to satisfy cpal's no-short-write contract. The driver thread therefore parks in source_eof for the lifetime of the process.

Because that thread holds a PlaybackStream clone (an Arc<InnerPlaybackStream>), the inner Arc count never drops to zero, InnerPlaybackStream::drop never fires, and the DeletePlaybackStream command it would otherwise queue is never sent. Each Stream created and dropped by the user thus leaks one server-side stream and one OS thread; long-running clients accumulate pactl list short sink-inputs entries until PulseAudio runs out of channels.

Queue a DeletePlaybackStream from Stream::drop so the reactor removes the stream state, drops the source's eof_tx channel, and wakes the driver thread with a cancellation error. now_or_never mirrors InnerPlaybackStream::drop and avoids blocking in Drop.

Fixes #1188.

[knz]

NB: cleanup is asynchronous (the delete is queued, not awaited), so a Drop returning does not guarantee the server-side stream is gone yet. That mirrors InnerPlaybackStream::drop and avoids block_on in destructors. If you prefer strict synchronous cleanup, we can use block_on instead.