A Flux model

[rossviljoen]

Adds a high level API for a model which can easily be used with Flux (although doesn't add a dependency on Flux itself).

Examples have been updated to use the model.

It should also be straightforward to adapt the model to support something like ParameterHandling.jl for use with general optimisers - this will probably be left to a future PR though.

[theogf]

Did not review everything in details but here are some comments.
Additionally, should we have some internal train and predict functions? We could add a dependency on Flux via Requires.jl?

[theogf]

Should we add a keyword argument to enforce the eltype and type of \mu and \Sigma?

[rossviljoen]

I added an eltype parameter, but I can't think of a reason why you wouldn't want to use Vector / Matrix, so maybe a type parameter isn't needed?

[theogf]

Well in the crazy case where someone would like to use GPUs, one would need pass CuArrays but I guess they would just have to make the effort of passing their own variational parameters.

[rossviljoen]

Since it's a functor, isn't it easiest to just use Flux.gpu for that after it's initialised?

[rossviljoen]

Did not review everything in details but here are some comments.
Additionally, should we have some internal train and predict functions? We could add a dependency on Flux via Requires.jl?

Training works with Flux's default training loop, so I'm not sure an internal train function would even make things much easier (besides setting defaults perhaps).

[rossviljoen]

Pretty much everything useful in this PR is in the regression/classification examples already, so closing