[Feature] Construct `mx.array` from `mps` and `cuda` arrays from other frameworks

[nbardy]

Describe the bug
dlpack doesn't work on mps tensors.

To Reproduce

Include code snippet

>>> import torch
>>> import mlx.core as mx
>>> mx.array(torch.tensor([1,2,3]).to("mps"))
Exception ignored in PyObject_HasAttrString(); consider using PyObject_HasAttrStringWithError(), PyObject_GetOptionalAttrString() or PyObject_GetAttrString():
Traceback (most recent call last):
  File "<python-input-7>", line 1, in <module>
RuntimeError: imag is not implemented for tensors with non-complex dtypes.
Exception ignored in PyObject_HasAttrString(); consider using PyObject_HasAttrStringWithError(), PyObject_GetOptionalAttrString() or PyObject_GetAttrString():
Traceback (most recent call last):
  File "<python-input-7>", line 1, in <module>
RuntimeError: imag is not implemented for tensors with non-complex dtypes.
Traceback (most recent call last):
  File "<python-input-7>", line 1, in <module>
    mx.array(torch.tensor([1,2,3]).to("mps"))
    ~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
ValueError: Invalid type torch.Tensor received in array initialization.
>>> mx.array(torch.tensor([1,2,3]).to("cpu"))
array([1, 2, 3], dtype=int64)
>>>

Expected behavior
A clear and concise description of what you expected to happen.

Desktop (please complete the following information):

• MacOS 15.5 (24F74)

>>> mx.__version__
'0.30.0'

Additional context
Add any other context about the problem here.

[awni]

Yes right now you can only construct MLX arrays from CPU tensors in other frameworks. It may be possible to change that to accept MPS / CUDA arrays.

[awni]

Regarding mps from PyTorch it seems like PyTorch does not support dlpack with MPS which is too bad. But until they do so we can't really accept MPS PyTorch tensors via that protocol.

import torch
t = torch.tensor([1,2,3]).to("mps")
t.__dlpack__()

RuntimeError: Cannot pack tensors on mps:0

See related discussion: wjakob/nanobind#1233 (comment)