add nrtr dml distill loss

[LDOUBLEV]

新增DistillationNRTRDMLLoss
原有DisstillationDMLLoss保持不变

        if self.multi_head:
              # for nrtr dml loss
              max_len = batch[3].max()
              tgt = batch[2][:, 1:2 + max_len]
              tgt = tgt.reshape([-1])
              non_pad_mask = paddle.not_equal(
                  tgt, paddle.zeros(
                      tgt.shape, dtype=tgt.dtype))
              loss = super().forward(out1[self.dis_head], out2[self.dis_head],
                                     non_pad_mask)

[paddle-bot]

Thanks for your contribution!

[tink2123]

LGTM