Allocating and reading 100 MiB at once for a remote file was a bit
excessive IMHO.

If there is a parent input, it cannot go away. There was no point in
reference counting.

This is arguably the only case where the callback made sense.

(The read callback is as suspicious and retained.)

In most cases, there is no or little room for optimizing this, and
the benefits are minimal. This provides a generic implementation so
that each stream filter does not need to reinvent the wheel.

Depending on device and audio format, AudioTrack can't hold an audio buffer big
enough.

This commit adds a circular buffer and a thread that will play the data to
AudioTrack coming from this circular buffer. This commit also decrease the
AudioTrack internal buffer size.