[ENH] Weighted win rates

[YannDubs]

adds the ability to use log probs for commuting the win rate, instead of a discrete win-rate

Todo:

• implement weighted win rate
• test backward compatibility
• add tests for weighted win-rate
• add a good prompt for weighted win rate
• add generalized metrics that work for logprob preference
• add ranking and correlation as a metric for the evaluator
• upload new baseline to HF
• compute minimal leaderboard with/without logprob and with/out cot