[runtime] refactor decoder, asr_model to support more platforms

[robin1001]

refactor the decoder and asr model to support more platforms in the future, such as onnx, j3, and so on.

TODO

• undefined reference error fix.
• asr model copy since there is inner state in asr_model now
• accuracy test, as the following, there is some bug in the old code when endpoint is triggerred. So the accuracy is slightlty different.

  // 3. Cache feature for next chunk
  if (state == DecodeState::kEndBatch) {
    // TODO(Binbin Zhang): Only deal the case when
    // chunk_feats.size() > cached_feature_size_ here, and it's consistent
    // with our current model, refine it later if we have new model or
    // new requirements
    CHECK(chunk_feats.size() >= cached_feature_size);
    cached_feature_.resize(cached_feature_size);
    for (int i = 0; i < cached_feature_size; ++i) {
      cached_feature_[i] = std::move(
          chunk_feats[chunk_feats.size() - cached_feature_size + i]);
    }
  }