Improved GPU Support

[nic-barbara]

Improve the speed of RENs and LBDNs evaluated on the GPU. At the moment, there is basic support allowing the the models to be trained/evaluated on the GPU, but much of the code has been optimised for the CPU.

[nic-barbara]

One particular area for improvement is the backwards pass of _N_lip and _N_gen in the case ny > nu. For some reason, there must be scalar operations happening under the hood when we write A / B instead of A * inv(B), which CUDA.jl is not a fan of. Investigate this further. This behaviour was first identified in 567f801.

Examples of where it is a problem:

• In norm_cayley when calculating B_T for DenseLBDN
• In _N_lip and _N_gen when calculating N for D22 term in LipschitzRENParams and GeneralRENParams

[nic-barbara]

The above #119 (comment) has been solved in d5bfd02, and was caused by adding the identity operator I to a CUDA array before back-propagating through A / (I + B).