the C version of that rather than float_to_int16.
Fixes output on ARM/VFP

Originally committed as revision 20192 to svn://svn.ffmpeg.org/ffmpeg/trunk