Reddit confessions dataset (#2045)

As mentioned in #2044 Augmented dataset for training safety models. 
- The dataset is made in the same format as prosocial-dialog for ease of
use.
- safety label confidence is added to enable users to filter data
according to a threshold of one's own preference.