all reduce per tensor dtype for embeddings and optimizer

all reduce per tensor dtype for embeddings and optimizer #3258