In oder to guarantee SIMDness between the two compilation steps we choose the
highest SIMD size that is optimized in the kernel for the first step.

wsfixed

Updated the list of supported input formats, added note about 16-bit
signed format (no fractional bits).

Make a note about 16-bit signed format - the function assumes that
values have no fractional bits (so 16-bit disparity from StereoBM
and StereoSGBM cannot be directly used!)

This is to achieve parity with the CPU equivalent.