Multi-GPU via threading instead of processes

[marius311]

WIP towards using one-GPU-per-thread instead of one-GPU-per-process. The big advantage is you don't need to launch multiple processes (which is slow and at least doubles the startup time) and memory can be shared between the GPUs using unified memory, instead of having to be serialized and distributed between the different processes (which necessarily passes through the CPU memory).

Right now this works:

tmap(collect(devices())) do dev
    device!(dev)
    for i=1:N
         gradient(ϕ -> norm(LenseFlow(ϕ)*f), ϕ)
    end
end

although MAP and sampling is still WIP.

Right now this needs CUDA 11.2 and my branch of CUDA.jl https://github.com/marius311/CUDA.jl/tree/no_gc_ctx_switch