Refactor mdspan cuda::std::__detectably_invalid

[fbusato]

Description

Replace old code in host/device/managed mdspan with https://nvidia.github.io/cccl/libcudacxx/extended_api/memory/is_pointer_accessible.html

[fbusato]

@davebayer @miscco I'm starting thinking that we are doing the wrong thing for host/device/managed.
There are two problems:

1. We cannot check the pointer validity in respect to the memory space for EVERY access. This involves multiple driver calls. This is too expensive even in debug mode.
2. We cannot check the device accessibility on the host because its access is always wrong on the host side.

Ideally, mdspan itself should check the pointer during the creating, not during run-time. Two potential solutions:

• cuda:std::mdpan checks __detectably_invalid.
• Modify the host/device/managed_mdspan constructs to check the memory space.

[miscco]

Why do we need that?

[fbusato]

it is very hard to reimplement all tests to work with the right accessor respect to the memory space.
I added this macro to disable the checks

[github-actions]

🥳 CI Workflow Results

🟩 Finished in 2h 10m: Pass: 100%/84 | Total: 1d 12h | Max: 1h 48m | Hits: 86%/197813

See results here.