use MLDataDevices.jl?

[CarloLucibello]

I think we should rely on MLDataDevices.jl for data movement functionality.
This would allow to reduce the codebase here and share functionality with the rest of the ecosystem.
Moreover, MLDataDevices has more functionalities implemented.

Unfortunately, the mapping among the two is not perfect, and the function get_device acquires a different meaning. This is a loose map between the two:

Flux	MLDataDevices
get_device(; [verbose])	gpu_device([device_id]; [force_gpu_usage])
get_device(backend::String, [idx])	cpu_device or gpu_device
gpu_backend!(backend::String)	gpu_backend!(backend::String)
supported_devices()	supported_gpu_backends()
X	get_device(x) -> dev::AbstractDevice
X	functional(x::AbstractDevice) or functional(::Type{<:AbstractDevice})
X	set_device!(T::Type{<:AbstractDevice}, dev_or_id)
X	loaded(x::AbstractDevice) or loaded(::Type{<:AbstractDevice})
X	default_device_rng(x::AbstractDevice)

We should be able to deprecate the old interface in favor of the new without a sudden breaking change.