Allow efficient=True when using CrossBatchMemory in DistributedLossWrapper

[KevinMusgrave]

When efficient=True:

• All embeddings should be added to each rank's CrossBatchMemory.embedding_memory
• Only the current rank's embeddings should be passed as the first argument to CrossBatchMemory.loss.forward()