In particular, these new functions don't need to be inside the face
recognition class.  So I moved them out.  I also fixed many incorrect
copy/pasted comments and clarified parts of the example code.

landmarking models.

point face landmark models.

Made object part annotations shown on the image_display scale relative to the size of their bounding box.

faster.  The CPU path is also a lot faster as well.

CPU cores.

sub_images of anything.  Also made the resize_image() functions a little more
flexible and compatible with the other image interpolation code in dlib.

about 2x faster.  The code is cleaner too.

stride values.  This lets you run the tensor resizing routine on subwindows in
a tensor.

because it's a much less stressing hand motion to hit END that i in most cases.

problem was that it wasn't counting detection boxes as false alarms if the
detection boxes had labels that never appeared in the truth dataset.

the label in the way you would expect.  In particular, loss_mmod_ is now a
multi-class loss and therefore capable of learning a detector that can output
detections with different labels.