do not calculate sum of error in every round of iteration;
instead the error will be summed every 2nd times or more, 
if the previous sum of error is too far away from threshold.

* use explicit extrapolation only for sub-matrixes
* use built-in interpolation for INTER_NEAREST mode

don't calc diff term if it is not used for epsilon criterion

incorrect types for template instantiation

Looks like we don't actually use them; when creating precompiled headers
with Visual C++, a dummy .cpp is created at build time.

ReST directive was used. Also fixed some other ReST directives that were
not correct and removed some warnings during buildbot checks.

- Used lunix style slashes - works on most systems
- Removed all trailing whitespaces

Turns out, you can't use preprocessor directives inside macro arguments.
Who'd have thought?

Now it allows choosing between arbitrary implementation variants.

+ interpolateFrames and createOpticalFlowNeedleMap

the output of BroxOpticalFlow differs a bit in CUDA 5.5

old version isn't compiled with CUDA 5.5
new version doesn't depend on functional.hpp

links with CUDA driver library only if we use video encoding/decoding