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

Provide the full API.

This corrects a major mistake from commit 724a59aa - there was one
small thing that commit was supposed to do, and it failed.
AV_TIME_BASE is not a seek flag.

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().