Add pictures for activation functions

[mcabbott]

Might be nice to have?

Not entirely sure how well these will render in different places. Also whether this package tests its own jldoctests.

[DhairyaLGandhi]

We could render them in the docs but large help sections that aren't critical can be annoying to use in the REPL.

[darsnack]

I like this idea! All the activation function docstrings are much shorter than typical docstrings e.g. Conv. I think having a unicode plot in REPL output would be nice feature (it would still be shorter than most docstrings).

[mcabbott]

I agree that huge docstrings you have to scroll through are a bit of a pain. I played with these a bit, and height=7 seems like a reasonable compromise size, at which you can roughly see what's going on, without taking the whole screen.

[ToucheSir]

How about under extended help? JuliaLang/julia#34226

[CarloLucibello]

I think it is very helpful to have these plots under standard help. I also like the fact that they will show up in the documentation site.
Also, the docstrings' length shouldn't be a concern here, they are very short and you consult them very rarely

[CarloLucibello]

needs a compat bound

[mcabbott]

I was following the other test deps (StableRNGs, Zygote) which don't have one.

Since there's no test dep on Documenter, I guess the docstrings aren't tested? Or maybe they are tested when building Flux's docs? Either way, there is then no need for a dependency here.

[mcabbott]

Alternatively, we can add documenter just to test the doctests. That's probably not a bad idea, better to find out sooner if your PR needs to tweak something. The code needed is something like:

if VERSION < v"1.6-" || VERSION > v"1.7-"
    @warn "skipping doctests, on Julia $VERSION"  
else
    using Documenter
    @testset "doctests" begin
        doctest(NNlib, manual=false)
    end
end

But perhaps I declare that out-of-scope for this PR.

[CarloLucibello]

sorry, got confused, doesn't need a compat entry (but i see you removed it entirely)

[DhairyaLGandhi]

Maybe it's better to not have to depend on a plotting library even if light, considering that this isn't expected to be used often. Not to mention, NNlib is supposed to be light. We can still do it in the docs, it's applicability is higher in the docs anyway.

[mcabbott]

Build failures are (at least) two different isapprox tests of convolutions, plus Zygote on Julia 1.8