simplify test machinery

[CarloLucibello]

This PR introduces a single function test_gradients that can be used across all possible differentiation scenarios (on cpu and any gpu backend, on layer or simple functions...).

[CarloLucibello]

@pxl-th I'm seeing something odd with the convolutions on AMDGPU
Could you test locally the following example? Looks like both assertions fail:

using Flux, AMDGPU, Zygote

m = Conv((2, 2), 3 => 4)
x = rand(Float32, 5, 5, 3, 2)

y, g = Zygote.withgradient(x -> sum(m(x)), x)

dev = gpu_device(force=true)    
m_gpu = m |> dev
x_gpu = x |> dev
y_gpu, g_gpu = Zygote.withgradient(x -> sum(m_gpu(x)), x_gpu)

@assert y_gpu ≈ y  atol=1e-4
@assert Array(g_gpu) ≈ g  atol=1e-4

[pxl-th]

@CarloLucibello something broke cpu-gpu transfer for convolutions.
MIOpen only supports cross-correlation, so we need to transpose filters when moving to gpu.

This is not executed anymore:

Flux.jl/ext/FluxAMDGPUExt/functor.jl

Line 89 in 09a16ee

function Adapt.adapt_structure(to::FluxAMDGPUAdaptor, m::CPU_CONV)

[CarloLucibello]

Ah right, that is when using MLDataDevices.gpu_device. Flux.gpu should work fine instead.
I'll try to fix it.

[CarloLucibello]

I investigated a bit and hooking up into MLDataDevices's data transfer mechanism is not straightforward. So I propose we merge this PR as it is and proceed with a fix later. The problem is not introduced by this PR in any case, it was pre-existing and now just exposed.

Can I get an approve?

[pxl-th]

Yes, I can take a look at it as well tomorrow, we can fix this is a separate PR.

[CarloLucibello]

I think the solution involves modifying this line
https://github.com/LuxDL/MLDataDevices.jl/blob/71ed455bb2a898a128b32aec0a67ba44fe8321d7/src/public.jl#L340
to use some custom and publicly exposed isleaf that we can hook into from Flux. Cc @avik-pal