`.to_numpy(copy=False)` Runtime Error if Device Memory

[ax3l]

We allow users to call .to_numpy(copy=False) on arbitrary memory.

This is fine even with pure GPU memory, to either:

• transport pointers around w/o access from the host or
• use as managed memory from the host (read/write)

For the situation that the pointer is in GPU memory and not managed, we should instead raise a runtime exception with the hint to use .to_numpy(copy=True), .to_cupy(copy=False) or activate managed memory.

We can use AMReX_GpuUtility.H for isManaged, isDevicePtr, isPinnedPtr helpers. It wraps cudaPointerGetAttributes and, once later supported, similar functions for HIP and SYCL.
https://docs.nvidia.com/cuda/cuda-runtime-api/group__CUDART__UNIFIED.html#group__CUDART__UNIFIED_1gd89830e17d399c064a2f3c3fa8bb4390

[ax3l]

These function can be quite expensive, so we should either use them sparsely or check alternatively the arenas if we know them.

These situations in AMReX can create this:

• in default Arena and managed allowed
• in explicit Managed Arena
• in explicit device arena and managed allowed

Then Arena has

    // isDeviceAccessible and isHostAccessible can both be true.                                                                                                             
    [[nodiscard]] virtual bool isDeviceAccessible () const;
    [[nodiscard]] virtual bool isHostAccessible () const;

    // Note that isManaged, isDevice and isPinned are mutually exclusive.                                                                                                    
    // For memory allocated by cudaMalloc* etc., one of them returns true.                                                                                                   
    // Otherwise, neither is true.                                                                                                                                           
    [[nodiscard]] virtual bool isManaged () const;
    [[nodiscard]] virtual bool isDevice () const;
    [[nodiscard]] virtual bool isPinned () const;

where isHostAccessible() is what we need.

[ax3l]

For MultiFab.array(mfi).to_numpy() we could go on the MultiFab level and add a:

MultiFab.to_numpy(mfi, ...)

function, that way we have still access to the Arena (or implement the more costly helper calls on the pointer from AMReX_GpuUtility.H above).

[ax3l]

The MPI-scope MF wrapper #370 uses self.array(mfi).to_numpy/self.array(mfi).to_cupy, based on GPU support (not arena). Can be generalized, but so far it will throw sensible runtime errors for the default arena.