[Not for merge] Add full-chunk mode CTC decoding for models from WeNet

[csukuangfj]

This PR shows that we can also use k2 for decoding with models from other frameworks.

For n-gram LM rescoring and attention decoder rescoring, it shares a lot of code with the following files and is very easy to implement.

• https://github.com/k2-fsa/k2/blob/v2.0-pre/k2/torch/bin/hlg_decode.cu
• https://github.com/k2-fsa/k2/blob/v2.0-pre/k2/torch/bin/ngram_lm_rescore.cu
• https://github.com/k2-fsa/k2/blob/v2.0-pre/k2/torch/bin/attention_rescore.cu

---

It's for demonstration only and not ready for merge.

If others have a need for full-chunk batch decoding on GPU with models from WeNet or other frameworks, we can support that.

[pzelasko]

Just for clarification: does "full-chunk decoding" mean "offline decoding"?

[csukuangfj]

Just for clarification: does "full-chunk decoding" mean "offline decoding"?

Yes, you are right. The name full-chunk decoding is from WeNet, which is misleading at the first sight.
I think it is equivalent to full utterance or full context.