Base code for SpeechModel

[ljj7975]

< summary >

• both mic and audio now generate same shape of input; dropped silence at the beginning of audio file
• part of layer definitions for SpeechModel has been implemented with tensorflow.js defined in SpeechModel.js
• base code for SpeechResModel.js
• implemented predict() for both models

< Things to Note >

• In layer definitions of SpeechModel.js, corresponding pytorch code has been copied and left as comment for future reference.
• config handling is not yet implemented. (currently using CNN_TRAD_POOL2 config for SpeechModel and RES8 config for SpeechResModel)
• ChannelLast (NHWC) data format is used because it is default for all layers.
• dnn layer for SpeechModel is not implemented yet
• implemented train function for testing compilation and execution of the model created; to be removed
• tensorflow.js generates error message WebGL: INVALID_ENUM: readPixels: invalid type. However, this seems to be irrelevant with the functionality and known to be expected (https://github.com/tensorflow/tfjs/issues/199)

[daemon]

Good work, looks fine to me. We can do a thorough refactor at the end.