rather than a compile time constant.  This also removes it from the input layer
interface since the DNN core infers its value at runtime, meaning users that
define their own input layers don't need to specify it anymore.

details objects can access each other.  Also added the input_layer() global
function for accessing the input layer specifically.

New versions of boost and cmake need to be told to look in boost's library folder to find boost's libraries
when building on windows.  Bizarre.

cudnnGetConvolutionBackwardFilterAlgorithm() to pick invalid algorithms,
resulting in cuDNN not working correctly.

batch normalization.  The running mean and variance need to be initialized when
calling cuDNN, even if the averaging factor is 1.  I don't think this was the
case prior to cuDNN v5, but it certainly is the case now.  This patch fixes
this.

as being signed rather than unsigned.

Added guards so dlib::async code isn't pulled in when using old versions of visual studio (since they don't support c++11).

in the file names to make it so there won't ever be any name conflicts
in the cropped images.

now the output images are always the same dimensions but the objects scale changes.

processes since sometimes stdin was closed in some environments.  Also cleaned
up the code a little bit.