feat: cuda-oxide explicit runge kutta solver

[martinjrobins]

• add an experimental solver using https://github.com/NVlabs/cuda-oxide called ExplicitRkOxide
• explicit runge kutta, hopefully can reuse most of the code from runge_kutta.rs if its suitable for cuda-oxide
• uses batching information from context (see feat: batching on linear algebra backends #298), constructs an array of OdeSolverState of length nbatch / warp_size, each warp operates on a single OdeSolverState (with nbatch_local = warp_size), so that each warp is an independent solver trajectory (i.e. each warp has its own dt=h)
• logically, this solver is one thread = one solve. But the threads in a warp cooperate to use a shared timestepper, so threads in a warp never diverge
• the array of OdeSolverState should probably live in shared memory, indexed by warp id.
• uses a new linear algebra backend with nbatch_local = warp_size, that is called from device code, so reductions like squared_norm() are done at the warp level (max reduction across the warp, then broadcast to all threads in the warp)

[martinjrobins]

The existing CUDA linear algebra backend is not suitable as it is meant to be called from host code, therefore we need new structs OxideVec, OxideMat etc that are meant to be called from device code

• the context for these has fixed nbatch equal to the warp size. The storage layout only needs to store warp_size vectors/matrices
• all operations (e.g. axpy) are just each thread operating on a single batch id
• all reductions (e.g. squared_norm()) are max over a warp and broadcast to the threads in that warp (so all threads in a warp will never diverge)

[martinjrobins]

presumably ExplicitRkOxide only implements OdeSolverMethod for Eqn::M=OxideMat

[martinjrobins]

I'm not sure how to make sure that these are not used from any solver other than ExplicitRkOxide...