IJulia fixes

[denizyuret]

• [x] Add convolutional net to the softmax IJulia. (done in dl-tutorial/cnn.ipynb)
• [x] Fix RNN to use the more efficient indexing approach in rnn IJulia and charlm example. (switched to cudnn interface)

[denizyuret]

• [ ] rnnlm shows that column major with input/output merge/split is more efficient in high vocab. (switched to cudnn)

[denizyuret]

• [ ] examples/overfitting/softmax.ipynb still missing convolutional net (could also add ladder networks etc to show learning efficiency).
• [ ] cnn notebook could use some hand-designed filters on real images.
• [ ] cnn notebook could use network visualizations of the type http://cs231n.github.io/understanding-cnn. http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture12.pdf
• [ ] cnn notebook needs architecture examples: http://cs231n.stanford.edu/slides/2017/cs231n_2017_lecture9.pdf
• [ ] cnn notebook needs backprop example.
• [ ] cnn notebook needs batchnorm example.
• [ ] the cnn chapter needs sections on visualization, architectures, batchnorm. also code and plots from the notebook.
• [ ] most rnn examples now use the new cudnn interface, check the notebooks.

[denizyuret]

Display iframe/html/javascript. Allow user to input a hand-drawn image.

https://github.com/JuliaGizmos/Interact.jl https://discourse.julialang.org/t/custom-widget-in-ijulia-interact-jl/3578 https://gist.github.com/SimonDanisch/b21e047d01500bac23df184c789a3dc6#file-draw-jl https://gist.github.com/denizyuret/9bdd25e122268ecf1bb73ac90c4ca2f3