All the files under src/lstm that start with tf.

It is useless. There is no way to train a model in TF that Tesseract can use for inference.

Can we finally remove the TF code from Tesseract?