Support for runtime CPU/CUDA selection

[kSkip]

This PR adds support for selecting either the CPU or CUDA implementation of the dnn module at runtime. As discussed in #1852, the #ifdef DLIB_USE_CUDA directives in tensor_tools.cpp just need to be replaced with a runtime branch like

if (use_cuda())
    cuda::function();
else
    cpu::function();

The thread local variable returned by use_cuda() is initialized to be consistent with the current behavior. If DLIB_USE_CUDA=True the function, use_cuda() will return true by default.

I found while making these changes that is was necessary to keep the preprocessor directives because the cuda namespace cannot be referenced when CUDA is not linked. Additionally, the pooling and convolution classes in tensor_tools.h have instances of both implementations as members. I couldn't find a cleaner way (with limited time) to handle those cases.

[davisking]

Nice, let me test this a bit and look it over.