Convert to 31 bit first, then right-shift with rounding to 16 bit.

Provide the full API.

The class does not have any virtual methods, and thus the (debug-only)
destructor does not need to be virtual.

These are the default values already.

Avoid multiplication.  This is a speedup of 20%.

Replaces its own weird buffering code.

Code simplification.

Not necessary anymore, because we enable this flag unconditionally
now.

Make sure glibc gives us all features.

This is nearly 4 times faster than the "portable" algorithm.

Use some odd number that will expose problems with remaining samples
after optimized vector operations.

Run each unit test only once.  Using CPPUNIT_TEST_SUITE_REGISTRATION
from within the header meant that each unit class was registered again
for each source file that includes the header.

The previous commit eliminated the need for that.

.. instead of reusing FloatToInteger<S24> and converting from S24 to
S32 in-place.

Refactor the conversion functions to classes and pass an instance to
the new function AllocateConvert().

-fixes SIGABRT when mpd is started from a previously paused state
-regression introduced by commit: f5a923b9:
'OutputAll: convert to class, move instance to class Partition'

-input_audio_format was previously declared using the static modifier
ensuring it was zero-initialised by default
-the current default-initialised input_audio_format contains garbage at
runtime which allows the AudioFormat.IsDefined() 'fail fast' test in
MultipleOutputs::Update to pass and the SIGABRT follows in
AudioOutput::Open when passed the invalid input_audio_format struct
-switching AudioFormat.IsDefined() for AudioFormat.IsValid() is an
alternative workaround

Fixes crash bug because mpd_connection_get_async() was called without
a connection.

-fixes regression introduced by:
 '8d6fedf8 [PATCH] Mixer: add class MixerListener'
-listener.OnMixerVolumeChanged() called GetVolume() which attempted to acquire
the lock but as per 'pa_threaded_mainloop_lock()' documentation:
This function may not be called inside the event loop thread. Events that are
dispatched from the event loop thread are executed with this lock held
-this patch seperates the underlying action of GetVolume() into a new
GetVolumeInternal() function, to be called only when the lock is already held, as
is the case for the listener.OnMixerVolumeChanged() call