Basic GPU Support

[nic-barbara]

Adding basic GPU support in accordance with #105.

[nic-barbara]

Main thing to change is removing scalar indexing when constructing parameterisations. Scalar indexing is not done on the GPU which can lead to very expensive data transfer back and forth between CPU and GPU.

[nic-barbara]

There's a scalar indexing warning popping up on the backwards passes of DiffREN for LipschitzRENParams and GeneralRENParams only. The rest are all fine. Look for common elements.

[nic-barbara]

This seemed like a good time to change the LBDN back-end a little bit. The number of hidden layers nh is now stored as an NTuple rather than a Vector. This means it's immutable, and won't get loaded onto the GPU when we don't actually want it to. I've still left the option of specifying nh with a vector in the constructor though, both for backwards compatibility and ease of use.

[nic-barbara]

There was a rather frustrating bug when evaluating the RENs on a GPU. In the src/Base/acyclic_ren_solver.jl, we used similar(z_eq, 1, size(b,2)) to initialise the product term Di_zi. On the CPU, similar() always initialises the array with extremely small values that are effectively zero. This is not the case on the GPU. On the GPU, similar() can initialise the arrays with basically anything, including NaN.

This was causing the REN to have different outputs when evaluated multiple times with the same inputs. Sometimes, the outputs would be NaN. Constructing Di_zi = typeof(b)(zeros(Float32, 1, size(b,2))) seems to work much better.

[jclinton830]

Looks good!