Implement Average Token Log Prob with Linear Scaling

[taylorlynn]

TBD I might need to make sub-issues. Basically, this ticket confirms that we are going with the linear scaling (i.e. mapping the average token logprob to its percentile within the 4500 examples I generated).