WebArena Verified

[NicolasAG]

Introduce the webarena_verified benchmark.

• tasks are registered with this template: webarena_verified.{intent_template_id}.{task_id}
• new WebArenaVerifiedTask class overrides the setup() function of GenericWebArenaTask to:
  • use the webarena_verified evaluator
  • load extra html headers if PW_EXTRA_HEADERS is set -- used to store secret keys to access self-hosted webarena instances
  • append a hint to the goal for the model to return the expected response format for the webarena_verified evaluator
• new WebArenaVerifiedEvaluator class that calls the webarena_verified.api.WebArenaVerifiedEvaluator from platform-labs-webarena-verified