Don't presume pointers location infers usability.

[robertmaynard]

Here is the results of looking at the cudaPointerGetAttributes of different allocation types on Grace + Hopper. Allocations of malloc are still usable on the GPU.

ccudaPointerGetAttributes attributes malloc ptr
  is_dev_ptr  -> 1
  is_host_ptr -> 1
  memory loc  -> unregistered

cudaPointerGetAttributes attributes cudaMalloc ptr
  is_dev_ptr  -> 1
  is_host_ptr -> 0
  memory loc  -> device

cudaPointerGetAttributes attributes cudaMallocManaged cudaMemAttachGlobal ptr
  is_dev_ptr  -> 1
  is_host_ptr -> 1
  memory loc  -> managed

[cjnolet]

@robertmaynard I just wanted to point out that with our mdspan APIs, we classify pointers as to their accessability instead of a direct classification of host vs device. That is, a device_mdspan only means the pointer can be read accessed from device and a host_mdspan means only that the pointer can be accessed from host. This way, we really didn't much need a managed_msdpan (though we create a managed_mdarray since that needs to know to use cudaMallocManaged.

[cjnolet]

/merge