`sym_sizes` implementation breaks on non-dynamic ops

[miladm]

🐛 Bug

I get the following error when integrating dynamic shape for sym_sizes in tensor_impl class. The error occurs on 47 of the seemingly non-dynamic tests.

Sample ERROR

======================================================================
ERROR: **test_put_xla_uint8** (__main__.TestTorchDeviceTypeXLA)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_device_type.py", line 390, in instantiated_test
    raise rte
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_device_type.py", line 377, in instantiated_test
    result = test(self, **param_kwargs)
  File "../test/test_torch.py", line 3326, in test_put
    out = torch.put(dest, idx, source) #, accumulate=accumulate)
RuntimeError: bad optional access

Looks like the sym_sizes initialization is missing to handle a corner case.

[miladm]

Turns out empty tensors needed caused crashes. This fix reduced the number of failures to 23.
I will share a PR shortly.

[miladm]

23 errors of similar signature remain. Looking into them. This might be due to how is_symbolic is initialized.

======================================================================
ERROR: test_index_reduce_reduce_amax_xla_bfloat16 (__main__.TestTorchDeviceTypeXLA)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_device_type.py", line 390, in instantiated_test
    raise rte
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_device_type.py", line 377, in instantiated_test
    result = test(self, **param_kwargs)
  File "/workspace/pytorch/xla/test/../../test/test_torch.py", line 2997, in test_index_reduce
    torch.maximum(expected[idx[i]], src[i], out=expected[idx[i]])
RuntimeError: bad optional access

[miladm]

When tensor_->GetIrValue().node->shape(0).dim() > 0 , tensor_->GetIrValue().node->shape(0).is_symbolic() can be empty. Added a check to guard against updating sym_sizes_ vector when is_symbolic returns empty.

[miladm]

I realized using xla::shape is the way to resolve the issue. These tests error are now resolved. Will reopen if needed.