add length penalty in reward + tracker avg positive & negative token length

[paulinebourigault]

This PR adds a length penalty option, along with the tracker for average positive and negative token length.

The length penalty is calculated using the formula: ((5 + length) / 6) ^ alpha.
This is based on the formula used in Google Neural Machine Translation paper [https://arxiv.org/pdf/1609.08144]

• When alpha < 0, shorter sequences get higher rewards
• When alpha > 0, longer sequences get higher rewards
• Length penalties are applied to the final reward scores, after the primary reward calculation

This is to follow my previous PR: tracker average token lengths for both positive and negative samples #1225

@vermouth1992 could you please review this?

[paulinebourigault]

Hi @hiyouga, what do you think about merging this?