Faster image pipelines

[lorenzoh]

Some low-hanging fruit that FastAI.jl can pick to become faster here:

• builtin presizing support
  • prototype at https://github.com/lorenzoh/DataBlockStorage.jl/
  • using JpegTurbo.jl's downsized loading so no storing to disk is needed (benchmark results against other backends JuliaIO/JpegTurbo.jl#15 (comment))
• use inplace operations
  • DataAugmentation.jl already supports inplace operations, FastAI.jl just needs to add a buffered encode! interface and use those paths
  • JpegTurbo.jl image loading could also be done inplace (Non-allocating image loading JuliaIO/JpegTurbo.jl#23)
  • the whole pipeline could become non-allocating
• use MLUtils.jl (MLUtils.jl transition #196)
  • new DataLoader seems more stable and possibly a little faster
• run projective transformations on GPU (Support for GPU-accelerated affine transformations DataAugmentation.jl#48)