Fix madspace_tests failures on the lemaitre4 HIP runner

[Copilot]

The lemaitre4 / madspace_tests job fails with 8 test failures on the HIP device. Two causes: tests convert GPU tensors with np.asarray, and a tar test writes archives with an out-of-range uid. Both are test-side problems. I couldn't run the tests in my environment, so the fix still needs a run on the HIP runner.

• test_beam_frame.py (7 failures)

  • np.asarray on a GPU torch tensor goes through DLPack and raises BufferError: Cannot export readonly array since signalling readonly is unsupported by DLPack.
  • All np.asarray calls in this file now use the to_numpy helper from madspace_testing.py, which copies to the host first.
  • to_numpy now wraps its argument in torch.as_tensor, so it also accepts plain numpy arrays.

• test_tar.py::test_read_tar_from_other_tool (1 failure)

  • tarfile raised ValueError: overflow in number field, because the runner account's uid doesn't fit the GNU and USTAR header fields.
  • tar.add now takes an anonymous filter that zeroes uid/gid and clears uname/gname:

def anonymous(info):
    info.uid = info.gid = 0
    info.uname = info.gname = ""
    return info

tar.add(p, arcname=p.name, filter=anonymous)