[XLA:CollectivePipeliner] Fix two issues:

[copybara-service[bot]]

[XLA:CollectivePipeliner] Fix two issues:

1. Accept transpose as a formatting op in ForwardSink.
2. Do not stop when a large collective was sunk in the previous iteration. Instead, delay sinking large collectives while sinking small collectives level by level and run an additional sinking iteration dedicated to large collectives at the end.