WordPiece, Sentencepiece, Refactor, Correct configurations

[nglehuy]

Features

• Another text_featurizer using wordpiece from tensorflow_text.FastWordpieceTokenizer
• Update SentencepieceFeaturizer using tensorflow_text.FastSentencepieceTokenizer
• Add tf_extract function in text_featurizer to support dataset on TPUs with use_tf: True option
• Add jit_compile option in model's compile (for faster fixed-shape training using XLA)
• Add gaussian weight noise in transducer decoder, wrapper function to apply and remove weight noises
• Add convolution blur pool

Fixes

• Refactor code (correct models, functions, unittest, configs, ...)
• Update ASRDatasets to support text-featurizer independent tfrecords (create tfrecords only with audio and transcript, instead of audio and indices)
• Replace some experimental options with their official supports
• Drop support for tensorflow < 2.8 (for older versions, please use TensorFlowASR ~= v1.x)
• Remove unused dependencies