Add proper token update to the db, in case generating a inference response fails

[jorgeantonio21]

Right now, we overestimate the number of tokens to be processed in the atoma state db, but if a response fails to be generated, we don't re-update the database to set it to 0.