ENH: allow running on torch.mps device, which does not have float64/complex128

[ev-br]

towards #431

Adding the following to __init__.py

import array_api_compat.torch as xp
xp_name = xp.__name__
xp.set_default_device("mps")

and running

$ ARRAY_API_TESTS_SKIP_DTYPES=uint32,uint64,uint16,float64,complex128  pytest array_api_tests/ -vs --skips-file=../array-api-compat/torch-xfails.txt --max-examples=10

generates a lot of failures, which, however, fall into three categories:

1. test_special_cases : a lot of edge cases are not implemented on an MPS device;
2. linear algebra: aten::linalg_eig' is not currently implemented, also aten::_linalg_svd.U, aten::linalg_qr.out
3. assorted edge cases on MPS with numerically large values: for instance, test_linspace fails on MPS and passes on CPU:

(Pdb) xp.linspace(start, stop, 100, dtype=xp.float32)[-1] - stop
tensor(-67108864., device='mps:0')
(Pdb) p np.linspace(start, stop, 100, dtype=np.float32)[-1] - stop
np.float32(0.0)
(Pdb) p np.linspace(start, stop, 100, dtype=np.float32, device="cpu")[-1] - stop
np.float32(0.0)
(Pdb) start, stop, num
(0.0, 1072614462193664.0, 100)