Fix: Fully align WOSAC metric calculation with original

[daphne-cornelisse]

Description

We observed a discrepancy between the WOSAC meta-scores from the original implementation and those produced by PufferDrive. This PR resolves these discrepancies by fixing the bugs below.

Bugs fixed

• Order of operations: We averaged log-likelihood metrics before exponentiating; WOSAC exponentiates first and then averages.

To do

• There is still a discrepancy of 0.03 between the SMART meta-score in Pufferdrive vs. the WOSAC leaderboard
• We need to evaluate a random agent in the leaderboard
• Update the baselines (clean small dataset, larger one)

Tests ran and results

TODO