Support for loading Keras models

[milancurcic]

This PR introduces support for reading simple Keras models from HDF5:

• New network constructor that creates a network from an HDF5 file
• Only input and dense layers are supported for now
• Only loading for inference is supported from now; i.e. optimizer data is ignored
• New example (mnist_from_keras) that demonstrates loading a pre-trained model and checking the accuracy of the output
• Introduces dependency on HDF5, h5fortran, and json-fortran
• More general download_and_unpack subroutine to download tar-balled data
• Placed nf_io_* modules under src/nf/io.
• Modernized CMake rules

TODO:

• [x] Add dependency rules to CMakeLists.txt (thanks @scivision)
• [x] Test with Intel Fortran

In support of #43.

[milancurcic]

@scivision, say I have an Intel build of HDF5 in /opt/hdf5-intel, with library and module files in lib/ and include/ relative to that path.

How do I pass this information to CMake so that it finds it?

I tried

FC=ifort cmake .. -DSERIAL=1 -DCMAKE_LIBRARY_PATH=/opt/hdf5-intel

which doesn't seem to do it.

I'd like to add this instruction to the README for people who want to run with Intel.

[milancurcic]

All good, I figured it out--it expects HDF5_ROOT to be set.