Coarse-To-Fine Tensor Trains for Compact Visual Representations