vertical scaler with accurate rounding, some people on doom9 can see +-1 errors

the +-1 issue is limited to >2tap vertical filters, so bilinear upscale was unaffected
the new code is sometime faster sometimes slower but the difference is significant (~20%) so its optional and enabled with arnd=1

Originally committed as revision 19177 to svn://svn.mplayerhq.hu/mplayer/trunk/libswscale