Support shallow fusion with RNNLM

[csukuangfj]

k2-fsa/sherpa-onnx#147
added shallow fusion with RNNLM.

We also need to support it in sherpa-ncnn.

The first step is to use k2-fsa/icefall#1050 as a reference to convert the model
to ncnn. Since the RNNLM uses only nn.Embedding, nn.Linear, nn.LSTM, it should be straightforward to
convert the model to ncnn via PNNX.

The next step is to use k2-fsa/sherpa-onnx#147
as a reference to add code to sherpa-ncnn. Since sherpa-ncnn does not support batch size > 1, the implementation should be much simpler.

Help from the community is highly appreciated. If you have any doubts, please leave a message and we can help you.