Fix BBEH generative evaluation

[lakshayxi]

Summary

• fix BBEH gold answers being treated as strings instead of choice lists
• evaluate BBEH as a generative task instead of using log-likelihood accuracy
• match the official BBEH answer normalization and equivalence rules in both LightEval and Inspect
• add focused regression tests for gold preservation, scoring, task wiring, and Inspect scoring

The existing BBEH task was added in #1124. This is a correctness follow-up to that implementation.

Refs #600.