Provide acoustic model score and LM score separately

[galv]

Instead of just providing a "score", see if we can provide am score and lm score separately.

Notes from a user:

From here,  https://huggingface.co/csukuangfj/icefall-asr-librispeech-conformer-ctc-jit-bpe-500-2021-11-09/blob/main/log/log-decode-2021-11-09-17-38-28#L19 , https://github.com/k2-fsa/icefall/blob/master/icefall/decode.py#L1072-L1075, we could find that even for the fixed unique paths, which means nbest oracle WER is fixed, to get the best WER, we have to jointly tune two of three weight (lm, attention weight in this k2 case.)