Allow CuArrays to initialise if there is no CUDA/GPU on the system

[tanhevg]

Currently an attempt to using CuArrays on a system that has no CUDA/GPU will result in an error. As a consequence, if one has a module that uses CuArrays, and wants to use some of its functions on a system without a GPU (e.g. debug a model training loop on a toy data set on a laptop on a plane), a separate .toml without CuArrays has to be created. This can be avoided by wrapping a few lines in the initialiser in a simple try/catch.

[maleadt]

Trying a different approach: #375 and JuliaGPU/CUDAapi.jl#84