Native GPU support

[MilesCranmer]

This PR adds native GPU support. This is a single CUDA kernel which evaluates an expression directly on the GPU!

TODO:

• [x] See whether CUDA.@captured helps at all
  • Nope...
• [ ] Explore whether manually manipulating CUDA streams can help with latency between launches (https://cuda.juliagpu.org/stable/lib/driver/#Stream-Management)
• [ ] See what I need to wrap with @sync
• [ ] Consider adding Optim support now or later

It works by launching a CUDA kernel to evaluate many independent nodes of a tree at once, from the leafs, upwards:

It also lets you evaluate multiple trees at once – which are dispatched to the GPU during the same kernel launches!

[MilesCranmer]

• [x] Actually this should probably use https://github.com/JuliaGPU/GPUArrays.jl instead and only test on CUDA (as an example array).