Can `RNG` be moved to `GPUArraysCore`?

From the definition of the struct, it seems like it should be as simple as copy-pasting it over. I can open a PR for this, but wanted to know if there was any reason not to do this previously.

`Metal` and `oneAPI` `default_rng` return `RNG` objects, so it is very useful (in the deep learning context) to define dispatches on `RNG` to initialize data directly on the device (for weight initialization, functions that need random numbers on runtime, etc.). To do this I end up defining weakdeps like:

1. https://github.com/LuxDL/WeightInitializers.jl/blob/main/ext/WeightInitializersGPUArraysExt.jl
2. https://github.com/LuxDL/MLDataDevices.jl/blob/main/ext/MLDataDevicesGPUArraysExt.jl

Moving to GPUArraysCore makes it significantly easier, since that is quite lightweight to not need extensions.

Start by locating the RNG struct definition and the Metal and oneAPI default_rng implementations mentioned in the issue, then compare the existing GPUArraysCore API. Confirm the package-boundary implications before moving it, and verify that device-specific dispatches can use the shared RNG without the cited weak dependency extensions.

Written by the indexing model from the issue text.