Core/AudioCommon: Feed bigger audio chunks to the output devices in case of underrun.
This should eliminate the crackling with alsa and pulseaudio backends and replace it
with much nicer pauses. This is only interesting for audio backends that do not
respect Mixer::GetNumSamples() and should not impact users able to run the DSPLLE at
The cost of this is an added LLE audio latency of about 0.06 s in the continuous
playback case. If that is too much, lower the low watermark.
Fixed r6235 for the Mac and Linux builds. Thanks, glennricster.