Created by: lcy-seso

1. remove n-gram language model, because it is repeated with word embedding in PaddleBook, and hsigmod, and NCE in Paddle Models.
2. refactor rnn language model and beam search.
3. rename the directory from language model into generate_sequence_by_rnn_lm to make it more accurate.