ilsvrc_2012_mean.npy has dims K x H x W.

Code written for the old D x D x K mean needs to be rewritten!

Do forward pass by prefilled or packaging input + output blobs and
returning a {output blob name: output list} dict.

Preserve the non-batch dimensions of blob arrays, even for singletons.

The forward() and backward() helpers take lists of ndarrays instead of a
single ndarray per blob, and lists of ndarrays are likewise returned.

Note that for output the blob array could actually be returned as a
single ndarray instead of a list.

Take blob args and give blob returns as single ndarrays instead of lists
of arrays.

Assign the net blobs and diffs as needed on the python side, which
reduces copies and simplifies the C++ side of the wrapper.

Thanks @longjon for the suggestion.

...and refer to inputs as inputs and not images since general vectors
and matrices are perfectly fine.

Don't run scripts in the module dir to avoid import collisions between
io and caffe.io.

For compositionality and expectations.