Add _aten_cdist Op

[vyom1611]

Based on issue: Op info test for cdist #7400

Right now the test is not passing locally:

The fail error in more detail:

E         AssertionError: Tensor-likes are not close!
E         
E         Mismatched elements: 150 / 150 (100.0%)
E         Greatest absolute difference: 2.0 at index (0, 0, 0) (up to 0.001 allowed)
E         Greatest relative difference: 1.0 at index (0, 0, 0) (up to 1e-05 allowed)
E         
E         To execute this test, run the following from the base repo dir:
E              python test/test_ops.py -k TestOpInfoCPU.test_reference_eager_cdist_cpu_float32
E         
E         This message can be suppressed by setting PYTORCH_PRINT_REPRO_ON_FAILURE=0

However the issue arises from fp arithmetic, as I ran an example to test the closeness of values:

Expected Result (PyTorch): tensor([[3.1193, 2.0959],
[2.7138, 3.8322],
[2.2830, 0.3791]])
Your Result (JAX): [[3.1192703 2.0958931]
[2.7138407 3.8321724]
[2.2830095 0.3791012]]

Can someone help me fix this precision error? Other than that this Op is ready to be merged I think.

[qihqi]

Hi, please rebase to HEAD to resolve conflicts. You can make the atol a bit larger for this op example: https://github.com/pytorch/xla/blob/master/experimental/torch_xla2/test/test_ops.py#L156

[ManfeiBai]

Thanks for the contribution, do we have any reference for others need to know why we need different implementation for these specific difference situtions?

[ManfeiBai]

Thanks, LGTM