ALSA: Avoid closing a device while the audio thread is still using it

[emezeske]

Changing the audio device can crash inside libasound, on setups using a large buffer at a low sample rate. A 4096 frame buffer at 8kHz is a 512ms read, and that is all it takes.

ALSAThread::close() waits 400ms for the audio thread to exit and then, if the thread still has i/o in progress, calls closeNow() on the devices from the calling thread to unstick it. That frees the snd_pcm_t while the audio thread may still be inside snd_pcm_readi or snd_pcm_writei, and the blocked thread then crashes inside libasound. The 400ms is shorter than a single read or write on plenty of perfectly valid configurations, so merely changing devices on such a setup is enough to free the handle out from under a thread that is not stuck at all and is just part way through one normal read.

This waits for the base timeout plus one i/o period, capped so that closing a genuinely wedged device stays responsive. A typical buffer size and sample rate is unaffected and still waits 400ms.