[cli] support on-the-fly training by loading pt model as nn.Module

[robin1001]

For example, we can use it as

class SpeechLLM(nn.Module):
    def __init__(self, encoder, llm):
        self.encoder = encoder
        self.llm = llm

    def forward(self, speech, text, speech_len, text_len):
        speech_emb = self.encoder.encoder.forward(speech, speech_lens)
        text_emb = self.llm.get_input_embeddings()(text)
        # ...


import wenet
speech_encoder = wenet.load_model_pt(xxx)
llm = Llama3()
speech = SpeechLLM(encoder=speech_encoder, llm=llm)