Code using Flux slow on GPU

There’s been a couple of PRs, but nothing GPU specific (https://github.com/JuliaLang/julia/pull/31563, https://github.com/JuliaLang/julia/pull/25558). Here’s something similar to Adapt.jl: https://github.com/JuliaGPU/GPUArrays.jl/issues/147#issuecomment-417255267. I know @keno had some thoughts about this too, but I don’t think he’s had the time to do anything with them.