Fix PaperBench zero prior-node limit

[sylvesterkaczmarek]

Summary

• make max_prior_nodes=0 return no prior rubric nodes
• preserve the existing unlimited behavior when the limit is None
• preserve positive limits as a tail slice of the available prior context

TaskNode.get_prior_nodes() currently applies required_nodes[-max_prior_nodes:] whenever a limit is provided. In Python, -0 is 0, so [-0:] is the same as [0:]. Explicitly requesting zero prior nodes therefore returns every prior node instead of none.

This can inject all available prior rubric context into SimpleJudge when a caller intentionally configures a zero-node context budget.

The fix handles zero explicitly while leaving positive and unlimited behavior unchanged.

Regression coverage verifies unlimited, zero, and positive-limit cases.