add Chinese distill-whisper fine-tuning results

[yuekaizhang]

In #1605, we fine-tuned whisper using 14k hours Chinese data. This PR added decoding results for distill-whisper fine-tuning experiment.

Instead of actually using distillation loss for training, the model structure and parameter initialization method from the distill-whisper paper (https://arxiv.org/abs/2311.00430) were adopted: only the first and last layers of the decoder were retained.

Accuracy:
Distill-whisper is slightly worse comparing with norm whisper.

Model	CER (Average SPEECH_IO_TEST_SET 01-26)	Training Set
whisper-large-ft-v1	4.32%	multi-hans-zh (about 14k hours)
whisper-large-ft-v1-distill	4.71%	multi-hans-zh (about 14k hours)

Speed:
Every decoding step could acclerate about 4x comparing with the original decoder.
Norm whisper: 32 decoder layers

Distill-whisper: 2 decoder layers

For a quick test: https://huggingface.co/yuekai/icefall_asr_multi-hans-zh_whisper/blob/main/test_model.py

[JinZr]

LGTM, thanks!