Determine which dim results in the fastest CCL operation

[tapspatel]

Currently, when we decompose all_reduce into reduce_scatter and all_gather, we arbitrarily select a tensor dimension to do the split along. We find which dimension evenly divides by the number of devices along the cluster_axis we are performing the CCL on, and if no such tensor dim exists, we throw an error. We should select the tensor dim that results in the best performance.