Exploiting Symmetric Temporally Sparse BPTT for Efficient RNN Training