Use tuples in KernelSum and KernelProduct; Make Zygote tests pass for KernelSum and KernelProduct; Improve doctring.

[sharanry]

#116 (comment)

@devmotion @theogf

[willtebbutt]

This is an excellent opportunity to provide an docstring demonstrating the various ways to construct these kernels / providing advice regarding when to use which options. In particular

• examples of how to construct a kernel with Tuple of kernels, kernels as separate args, and a vector of kernels.
• something like: "you should use a Tuple if there aren't many components, but a vector if there are lots"

Preferably the examples would be included as doctests to prevent documentation-regression.

[devmotion]

I'm wondering if we should remove the weights in KernelSum. They seem a bit redundant since every kernel can be scaled anyways.

[theogf]

What about what @devmotion was saying about for large collections of kernels, tuples would be inefficient?

[theogf]

I agree for the weights problem we should remove them

[theogf]

Maybe it's also the good place to check that Zygote tests are passing!
Edit: I tried the Zygote tests and they seem to pass now.

[sharanry]

There seems to be a dependency error with Julia 1.3 build. Any idea what this is about?
https://travis-ci.com/github/JuliaGaussianProcesses/KernelFunctions.jl/jobs/368383076#L341

[devmotion]

Maybe it was due to https://discourse.julialang.org/t/pkg-downtime-incident/44288

[willtebbutt]

Lightly reviewed for style / docstrings.

[devmotion]

LGTM, just a final docstring suggestion.

[willtebbutt]

This is excellent. Really glad to have it in!

Removing the weight fields is breaking, so could you please bump the minor-version and release :)