Road to TrixiParticles.jl on GPUs

[efaulhaber]

• Make array types in structs generic. Change types like Array{ELTYPE, 2} to type parameters.
  Make array types in structs generic #486
• Recursively convert (adapt) Semidiscretization object to GPU types with Adapt.jl.
  Recursively convert (adapt) Semidiscretization object to GPU types with Adapt.jl #492
• Add type GPUSystem to determine if a system is stored on the CPU or GPU #532
• Adapt some code to GPUs #533
• Minimally invasive working GPU implementation with KernelAbstractions.jl without neighborhood search.
  Two alternatives: Either Minimally invasive GPU implementation with KernelAbstractions.jl without neighborhood search #493 using plain kernels for everything or Rewrite @threaded macro to work with GPUs #534 hiding the kernels in the @threaded macro.
• Rework neighborhood search to be more efficient and GPU-compatible. Roadmap for neighborhood search implementations PointNeighbors.jl#26
• Make SolutionSavingCallback work on GPUs (Make SolutionSavingCallback work on GPUs #579).
• Make TLSPH work on GPUs.
• Benchmark, profile and optimize interactions on GPUs.
• Avoid synchronization. For example, fluid-solid and solid-solid interactions have no race conditions, so all *-solid interaction kernels can be spawned on the GPU without synchronization. The same holds for all *-fluid and *-boundary interactions. However, this should happen after optimization, since we will lose timers for each individual interaction. Also, we will only see a benefit for multi-physics simulations, as a simple fluid-boundary simulation can't run any two non-trivial interactions without synchronization. Note that a single fluid-* kernel might also be able to improve performance.