allow export of onnx-streaming-models with other than 80dim input features

[KarelVesely84]

• i could also add surfacing of --feature-dim into train.py,
• and also feature extraction had to become configurable in compute_fbank_librispeech.py and streaming_decode.py, if you are interested in having the recipe with configurable feature extraction...

[csukuangfj]

Thanks! We can change other files if needed.
(The latest ./zipformer folder is more likely to be used than other folders, so it is fine to change only ./zipformer at present.)