Make CUDA an optional dependency in Julia 1.9

[mohamed82008]

Motivation and description

CUDA and GPUCompiler are heavy dependencies and are often not needed in non-CUDA systems.

Possible Implementation

The new pkg extensions in Julia1.9 can be used.

[ToucheSir]

The only code which relies on CUDA.jl is https://github.com/FluxML/Metalhead.jl/blob/4e5b8f16964468518eeb6eb8d7e5f85af4ecf959/src/layers/drop.jl#L48C1-L52, so extracting that into an extension should be sufficient.

[theabhirath]

With all the different GPU backend work in Flux recently to allow for Metal and AMD, is there a way to write this code so that it just works for all of them? Also something to fix is that Metalhead uses a private function from Flux here (Flux.isactive, which I believe may have been renamed upstream). GPU testing is also an issue...highlighting all these here because they just occurred to me 😅

[ToucheSir]

For RNGs? Probably not, because I don't think the GPU packages have a common interface or supertype for their RNG types.

Some sort of getter for the train/test status of a layer would be a nice addition to Flux. I think we'd just have to bikeshed the name. _isactive as used internally does something other than that.