Using float instead of double for FHOG computations

MSVC x64 SSE2 automatical support

Python/Simple object detector trainer. Added upsample_limit option

introduce a configurable threshold for rendering HOGs

Fixed a minor typo in the spec.

Support use of dlib a cross-compiled ExternalProject

the former works with a wider range of tools in the STL.

This follows convention which simplifies using dlib as a CMake
ExternalProject.

since sometimes gfortran isn't available and openblas should already be linked
to gfortran if it was needed.

orthogonalize does not work with fixed-sized matrices