Refactor to allow multiple solvers and compute backends

[lrntct]

The goal is to allow the following:

• Current partial inertia solver implemented with Cython
• Future alternative backend (for example GPU implementation of the Surface Flow components #11 )
• Future SWE solver (Add a Full SWE solver #14 )

The architecture would be:

• Simulation does not change
• SurfaceFlowModel represents the equation family
• ComputeBackend performs low-level array kernels and reductions
• ArrayStore owns array allocation, padding, swapping, and host/device synchronization
• RasterDomain remains the domain-facing object, but becomes a thin facade over ArrayStore instead of directly owning raw NumPy arrays forever
• rastermetrics.pyx is split between ArrayStore and ComputeBackend