[NVIDIA TF] CollectiveAllToAllV2 + NCCL all-to-all fixes

[trevor-m]

This PR fixes a few small bugs with CollectiveAllToAllV2 and NCCL all-to-all:

• NCCL does not support in-place all-to-all, so don't reuse the input tensor in CollectiveAllToAllV2. We could potentially keep forward_input_or_allocate_output and allocate a temporary buffer inside the NCCL backend only when the input is reused. This would allow other backends to still perform an in-place all-to-all. It looks like the ring algorithm uses temporary output buffers anyway.
• Update calls to ncclSend and ncclRecv to follow the mapping by using global_rank. The TF runtime can sometimes reorder the GPUs to optimize the ring algorithms. While this PR fixes nccl all-to-all to follow the ordering, the ring algorithm is also affected and gives incorrect results.
• VLOG outputs in nccl_manager.cc would cause a crash due to interpretting the data buffer as a string.

[trevor-m]

cc @rainwoodman

[chsigg]

Thank you!