Add recipe for fine-tuning Zipformer with adapter

[marcoyang1998]

This PR supports fine-tuning a Zipformer transducer model with adapters (https://arxiv.org/pdf/1902.00751.pdf). The original model parameters are untouched and we only update the parameters in the adapters on the new domain.

The following table shows the WERs of different models on GigaSpeech test sets in a domain adaptation setting.

Model config	trainable params	dev	test
pretrained (without fine-tuning)	-	20.06	19.27
fine-tune whole model	66M	13.56	13.64
fine-tune with adapters	1.5M	15.05	15.18

By only introducing around 2% trainable parameters, the model with adapters achieves much lower WERs than the baseline model without adaptation.

[marcoyang1998]

After training, the model can be tested on both the original and the new domain.

To deactivate the adapters after training, set --use-adapters False. The performance on the original domain should be untouched.