For a 0.5.0 release

[ablaom]

#248
#251

For the release notes:

• (new model) Add NeuralNetworkBinaryClasssifier, an optimised form of NeuralNetworkClassifier for the special case of two target classes. Use Flux.σ instead of softmax for the default finaliser (add constructor for binary classifiers #248)
• (internals) Switch from implicit to explicit differentiation (Omnibus PR, including switch to explicit style differentiation #251)
• (breaking) Use optimisers from Optimisers.jl instead of Flux.jl (Omnibus PR, including switch to explicit style differentiation #251). Note that the new optimisers are immutable.
• (RNG changes.) Change the default value of the model field rng fromRandom.GLOBAL_RNG to Random.default_rng(). Change the seeded RNG, obtained by specifying an integer value for rng, from MersenneTwister to Xoshiro (Omnibus PR, including switch to explicit style differentiation #251)
• (RNG changes.) Update the Short builder so that the rng argument of build(::Short, rng, ...)
  is passed on to the Dropout layer, as these layers now support this on a GPU, at
  least for rng=Random.default_rng() (Omnibus PR, including switch to explicit style differentiation #251)
• (weakly breaking) Change the implementation of L1/L2 regularization from explicit loss penalization to weight/sign decay (internally chained with the user-specified optimiser). The only breakage for users is that the losses reported in the history will no longer be penalized, because the penalty is not explicitly computed (Omnibus PR, including switch to explicit style differentiation #251)