`test_diff_xla_bool` takes hours when adding dynamism to `XLATensorImpl::sym_sizes_custom()`

[miladm]

🐛 Bug

When enabling dynamism on sum_sizes_custom(), the following test takes hours to complete.

test_diff_xla_bool (__main__.TestTorchDeviceTypeXLA) ...

[miladm]

Turns out this test has the following call. To address this issue, we need to materializing the size() via calling _test_diff_numpy()

...
       for dim in dims if dims else range(t.dim()):
            prepend = t.narrow(dim, 0, 1)
            append = t.narrow(dim, 0, 1)
            np_t = to_np(t)

            # test when no prepend and append
            for n in range(t.size(dim)):                         <-- materialize size()
                actual = torch.diff(t, dim=dim, n=n)
...

Without this implementation, t.size(dim) returns a random number, and t.size() returns torch.Size([])

[miladm]

We do need to materialize the tensor shape for dynamic ops. In this case, since the tensor is static, we don't have to include that feature to resolve the failure here. This issue was resolved after I used xla::shape to set up sym_sizes_ as opposed to torch::lazy::Shape. Closing the issue. Will reopen if needed.