Alphablending optimizations:
- validate palette alpha values in overlay manager
  (one check / palette index) instead of
  checking every alpha value twice for every
  blended pixel in every frame
- remove unneeded calculations
- approximiate expensive integer divisions with 
  multiplication and shift
- simplify if(...) 's by pre-calculating values 
  that are needed later anyway
- remove local variables to save registers

This results significant savings in CPU time 
when blending large semi-transparent overlays.
Quick test blending VDR main menu to
720x576 mpeg video, Core2 E6600:

xine-lib 1.1.6, de-interlacing (-> YUY2 Xv)
original:  video 32%, video+OSD 61% (OSD 31%)
optimized: video 32%, video+OSD 52% (OSD 22%)
---------------------------------------------
speedup               total     22% (OSD 33%)

xine-lib 1.1.6, YV12 Xv:
original:   video 5%  video+OSD 18% (OSD 13%)
optimized:  video 5%  video+OSD 14% (OSD 9%)
---------------------------------------------
speedup               total     22% (OSD 40%)