gather: improve error handling and error messages.

[ysiraichi]

This PR refactors the tensor_methods::gather implementation by improving its error message, and returning a status type value.

Key Changes:

• Make tensor_methods::gather return StatusOr<absl_nonnull XLATensorPtr>
• Improve error message on incompatible tensor shapes

Example:

input = torch.arange(16, device=torch_xla.device()).view(2, 2, 2, 2)
index = torch.randint(0, 2, (5, 2, 5, 2), device=torch_xla.device())
dim = 1
torch.gather(input, dim, index)

Before:

Traceback (most recent call last):
  File "example.py", line 28, in <module>
    torch.gather(input, dim, index)
RuntimeError: Check failed: index->size(dim) <= input->size(dim) (5 vs. 2) (at torch_xla/csrc/tensor_methods.cpp:1826)

Exception raised from operator& at torch_xla/csrc/runtime/tf_logging.cpp:26 (most recent call first):

After:

Traceback (most recent call last):
  File "example.py", line 28, in <module>
    torch.gather(input, dim, index)
RuntimeError: gather(): expected sizes of index [5, 2, 5, 2] to be smaller or equal those of input [2, 2, 2, 2] on all dimensions, except on dimension 1. However, that's not true on dimensions [0, 2].

Status Propagation Trace:
    From: CheckDimensionsAreCompatible at torch_xla/csrc/tensor_methods.cpp:440 (error: gather(): expected sizes of index [5, 2, 5, 2] to be smaller or equal those of input [2, 2, 2, 2] on all dimensions, except on dimension 1. However, that's not true on dimensions [0, 2].)
    From: gather at torch_xla/csrc/tensor_methods.cpp:1865

Exception raised from MaybeThrow at torch_xla/csrc/status.cpp:121 (most recent call first):

[zhanyong-wan]

Thanks!