This new version:
1.  Works on 24-bit and 32-bit input, not just 32-bit.
2.  Doesn't try to run on 16-bit or 8-bit, instead of outright crashing.
3.  Does proper alpha-blending, so cursor shadows look correct.
4.  Doesn't swap R and B.

Mostly fixes issue 1997.
Fixes issue 2056.

Originally committed as revision 25690 to svn://svn.ffmpeg.org/ffmpeg/trunk