Create ImageDescriptor

[etrain]

Both SIFT and Daisy feature extractors output a DenseMatrix of a special shape that contain a bag of image descriptors. These image descriptors also have metadata with them (e.g. original positions) which would allow for things like spatial pooling that we currently discard.

We should create a class for ImageDescriptors that captures this metadata. One downside is that currently things like PCA, GMM, and FV rely on the results of SIFT or Daisy being DenseMatrix, and we'd have to rethink the types there to handle this change in a clean way.