Fast Lexically Constrained Decoding with Dynamic Beam Allocation for Neural Machine Translation