Optional silence trimming during inference and find_endpoint() fix

[george-roussos]

This commit:

1. grabs the threshold db for find_endpoint() from the config.json file and
2. makes silence trimming optional during inference, by checking whether do_trim_silence is set to True or False. I am understanding the value is mostly there for training and removing lead/trail silences during preprocessing, but I feel it fits here also (we can anyway change this before inference).

For find_endpoint(), I am a bit skeptical about whether that value should be set to e.g. -20dB more than what trim_db is (to make it more robust). Please tell me what you think about this idea, or whether it should be the same.

[CLAassistant]

All committers have signed the CLA.