Add retries for empty model responses

[linear]

Empty model responses are occurring in roughly 1-3% of evaluations and are affecting performance.

This appears to affect multiple OpenRouter models. The work is to investigate and add retries in verifiers for empty model responses.

Relevant examples: evaluation 1, evaluation 2.

[NoahBPeterson]

GPT-OSS-120b is failing on >80% of rollouts for the primeintellect/mini-swe-agent-plus environment due to tool call errors, which are getting parsed as no tool calls, and due to #1537, terminating the rollout as only the reasoning and not content or tool_calls are populated. This would very likely resolve my issue.