Automatic GPUified linear solves

[ChrisRackauckas]

https://medium.com/datathings/benchmarking-blas-libraries-b57fb1c6dc7 shows that at around 100x100 systems it can be beneficial to round-trip through a GPU. We can create a linear solver routine which does just that: sends things to the GPU, does the factorization, then does the linsolve by sending to the GPU, \, then send back.

For sparse, it's less clear if that's a good idea ever: https://arxiv.org/pdf/1608.00636.pdf

The default linear solver can make use of this by querying for a GPU through https://github.com/JuliaGPU/CUDAapi.jl .

[ChrisRackauckas]

Non-default solved by SciML/DiffEqBase.jl#269

[ChrisRackauckas]

Default handling is solved by SciML/DiffEqBase.jl#273

[ChrisRackauckas]

LinearSolve.jl has this.