Fix type promotion for pow.

[ysiraichi]

Fix: #6521

This PR fixes how PyTorch/XLA promotes types for pow function. It does so by resolving the resulting type of the operation by calling at::result_type (similar to how add is implemented in PyTorch/XLA). This aligns the implementation with PyTorch's. See the example below:

def foo(a):
    return torch.pow(a, 3.0)

a = torch.arange(10).to(torch.bfloat16)

print(foo(a))
# tensor([  0.,   1.,   8.,  27.,  64., 125., 216., 344., 512., 728.],
#        dtype=torch.bfloat16)

print(foo(a.to(xm.xla_device())))
# tensor([  0.,   1.,   8.,  27.,  64., 125., 216., 343., 512., 729.],
#        device='xla:0')

cc @miladm @JackCaoG

[ysiraichi]

I will add a test case for this, later.

[JackCaoG]

I think @qihqi added this downcasting rule and you removed it in this pr.

[ysiraichi]

I don't think we need this anymore, since at::result_type resolves to the appropriate type, as we can observe with the tests.

[JackCaoG]

can you add a test to test/test_operation.py for the result type?

[JackCaoG]

not sure why CI is not running...

[ysiraichi]

Maybe due to merge conflicts? I will try rebasing.