ENH: Add basic dlpack testing

[ev-br]

Add a basic set of dlpack tests. The tests here are "boring":

• no cross-library tests, only test from_dlpack for a single array library
• only test copy=False on the same device
• no tests for __dlpack__ yet several XXX comments to mark where I've no idea.

[ev-br]

Issues with these "trivial" tests:

1. torch has a weird device(type="meta") which is incompatible with DLpack. If anything, the following should raise a BufferError IIUC, not a ValueError

In [8]: device = torch.device(type="meta")

In [9]: x = torch.empty(3, device=device)

In [10]: torch.from_dlpack(x)
---------------------------------------------------------------------------
ValueError                                Traceback (most recent call last)
Cell In[10], line 1
----> 1 torch.from_dlpack(x)
...
ValueError: Unknown device type meta for Dlpack

2. jax fails with cupy=True on CUDA where it works on CPU:

In [8]: jnp.from_dlpack(jnp.ones(3), copy=True)
---------------------------------------------------------------------------
JaxRuntimeError                           Traceback (most recent call last)
Cell In[8], line 1
----> 1 jnp.from_dlpack(jnp.ones(3), copy=True)

...

JaxRuntimeError: UNIMPLEMENTED: PJRT C API does not support HostBufferSemantics other than HostBufferSemantics::kImmutableOnlyDuringCall, HostBufferSemantics::kImmutableZeroCopy and HostBufferSemantics::kImmutableUntilTransferCompletes

3. cupy==14.0.1 does implement copy or device arguments:

>>> x =  cupy.array([], dtype=float)
>>> cupy.from_dlpack(x, copy=True)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "cupy/_core/dlpack.pyx", line 568, in cupy._core.dlpack.from_dlpack
  File "cupy/_core/core.pyx", line 355, in cupy._core.core._ndarray_base.__dlpack__
BufferError: copy=True only supported for copy to CPU.

>>> cupy.from_dlpack(x, device=x.device)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "cupy/_core/dlpack.pyx", line 554, in cupy._core.dlpack.from_dlpack
NotImplementedError: from_dlpack() does not support device yet.

>>> cupy.__version__
'14.0.1'

4. numpy cannot consume a capsule created by __dlpack__, while torch happily does. Which one is correct?

>>> import numpy as np
>>> import torch
>>> x = np.ones(3)
>>> capsule = x.__dlpack__()
>>> torch.from_dlpack(capsule)             # works fine
tensor([1., 1., 1.], dtype=torch.float64)
>>> np.from_dlpack(capsule)                 # ouch
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
AttributeError: 'PyCapsule' object has no attribute '__dlpack__'

[ev-br]

Larger issues:

5. How to tell if a device is dlpack-compatible. The only way I see now is to try-except call from_dlpack and catch a BufferError.

6. How to tell if a copy=False device transfer is meant to work, given source_device and target_device. A first reaction is to create a dummy array on each device and check their __dlpack_device__ values. If they match, it should work, ok. But then what about CUDA vs CUDA_MANAGED? CPU vs CPU_PINNED? Other values? It feels like there needs to be something in the inspection capabilities to tell it.

7. How to tell a supported max_version in __dlpack__. Should max_version=(-111, 42) raise? What about reasonable version values: max_version=(3, 0) if a library only supports (2,0).