Change tokenizer to make it work with both Go and Python

[mostafa]

The current tokenizer uses a deprecated tf.keras.preprocessing.text.Tokenizer, which is only available to Python. This means that the tokenizer should be made available to the plugin using a REST API, which adds overhead. The tokenizer should be replaced by something like sentencepiece that is also available to Go: go-sentencepiece-encoder.

Related

• Optimize preprocessing steps #4