Destroy clients before trying to remove nodes. If the clients are
still connected they might try to react to the removed nodes.

Keep track of global mix port.
Calculate the get_buffer function beforehand.

This ensures we first finish all pending operations and then get
notifications of the new events. Brings pactl subscribe in line
with what it does on real PulseAudio.

First sync the globals and if no global created new pending operations,
run through the operations. This ensures no operation is run before
all globals have data.

When a stream does not specify ADJUST_LATENCY, use some slightly
different metrics to implement the readable_size. We want the app
to fill as much space as possible. Makes mplayer work and paplay
use the right amount of buffering.

Some drivers need a mmaped buffer...

Some drivers seem to only use 2 buffers when asked for 16 buffers. We
can see this because the buffer is already queued. In this case, just
use the 2 buffers.

See #294

Normalize the complete matrix with the same coefficient

Try to limit the amount of data we can write. Fixes chrome playback.

Only try to release the device when we have a proxy.

Fixes #310

The queued bytes are already included in the read/write index.

When we have accurate enough hardware volume, set the software volume to
1.0. Also fixes the issue of going to 0 volume and then staying silent
until we raise the volume above 1.0.

The Off profile is always available but we should only use it as the
ultimate fallback when all other profiles are NO.

Fixes #309

If we can't use EXPBUF, just put the fd and mapoffset in the buffer
memory. This way we can pass the fd to the client and let it mmap .

If EXPBUF returns and error, try to mmap instead.

Fixes #294

If the port was out of buffers and we can't recycle a buffer, return
-EPIPE.

So that we don't have to depend on the header file.

Fixes #306

Don't round up, we give to the app what we get and use the fragsize
always when recording.

Seems to work well with vlc and firefox other apps now.
paplay without ADJUST_LATENCY reports wrong latency still.

The flag means that the process function might not complete
synchronously.

We can use this knowledge to improve the adapter. In sync mode we
can pull before scheduling the converter. In async mode we need to
schedule the follower after the converter to get the data ready for
the next iteration.

We can also improve the stream when it is operating async to schedule
a process call when a new buffer is ready to be filled.

This reduces a cycle latency from alsa and improves latency in
pulse.

firefox likes that better

Use tlength as the latency.
Round attributes to frame size
Only use the delay in get_time().
Tweak queued, writable and required bytes for callbacks.

Use tlength as the latency.
Round attributes to frame size
Only use the delay in get_time().
Tweak queued, writable and required bytes for callbacks.