#sdy For a future all-to-all from source dimension d1 to target dimension d2, slice available axes on d1 instead d2 to avoid a redundant collective-permute if certain conditions are met.

#sdy For a future all-to-all from source dimension d1 to target dimension d2, slice available axes on d1 instead d2 to avoid a redundant collective-permute if certain conditions are met. #1140