Optimize collective_permute source_target_pairs size == 0 lowering

[tapspatel]

Similar to this issue in all_gather where we replace replica_groups size = 1 with a noop, for the case of collective_permute, when we get a source_target_pairs with size = 0, we need to also convert that into a nop, as that means no tensor data is moving from any device, and converting it into a noop won't change the semantic meaning of the graph.

This is blocked by this PR: #2405