all reduce per tensor dtype for embeddings and optimizer (#2778)

all reduce per tensor dtype for embeddings and optimizer (#2778) #3290