#18 Smaller filter for upconversion

[Thomas Davies]

I've changed upconvert.{cpp,h} to use a smaller filter
(6 taps instead of 12) for 1/2-pel upconversion, for
use in sub-pixel motion compensation and estimation.
The new tap values are:

158
-37
7

extended symmetrically. This reduces upconversion time
by about 35%. The filter is also short enough that
motion compensation could be refactored to do the
half-pel interpolation on the fly and eliminate
explicit upconversion altogether: this might be faster.

[Thomas Davies]