not destruct and recreate themselves when you switch devices.  Instead, they
keep a table of context objects, for each thread and device, reusing as necessary.

This prevents churn in the context objects when you are switching back and
forth between devices inside a single thread.

order if that's the memory layout of the argument.

with a tensor.

written to with assign_image().  However, you get an error if you try to
assign an image with a different size.

compilation errors when they get pulled into the scope of template
metaprogramming expressions.

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.