Basic GPU Support

[nic-barbara]

We need to add in GPU support for all REN and LBDN models.

• Check typing on all structs is compatible with CUDA arrays and allows transferring to GPU
• Write a test to send all REN models/parameterisations to GPU
• Debug issues arising from this test
• Repeat for LBDN
• Repeat for SandwichFC
• Test on MNIST for LBDN
• Test on observer for REN

Need to check whether having GPU in the CI is actually possible too.

[nic-barbara]

No need to write gpu and cpu methods for all our types anymore. As long as they are all functors (specified with @functor) then this will generalise with Flux.jl, as outlined here.

[nic-barbara]

NOTE: Using @functor for gpu specification:

Summary: whatever you specify in the brackets of @functor MyType (...) is what gets loaded onto the GPU.

---

Writing

@functor ExplicitRENParams (A, )

means that only explicit.A gets loaded onto the GPU when you type explicit |> gpu. If we instead write

@functor ExplicitRENParams
trainable(m::ExplicitRENParams) = (; )

then all fields of explicit get loaded to the GPU, but Flux.trainable still returns an empty named tuple (so they won't be trained on). This is the behaviour we want.

Note that we use @functor ... for the two base parameter types DirectRENParams and DirectLBDNParams, so this is fine.