fix: Handle RULER rewards when all trajectories are identical

[vivekkalyan]

Due to the way we efficiently deal with common substring, there is an edge case where RULER would send empty trajectories to the judge when all trajectories in a group were identical. This caused scores of 0 for that group, which while it doesn't affect the advantage, might be alarming.

This PR detects this edge case and sends only 1 trajectory to the judge and duplicates the result to all trajectories. We still end up having the same score across all trajectories, just now it's a number chosen between 0-1 by the judge.

To test this case:

    message_lists = [
        [
            {"role": "system", "content": "You are helpful."},
            {"role": "user", "content": "What is 2+2?"},
            {"role": "assistant", "content": "4"},
        ],
        [
            {"role": "system", "content": "You are helpful."},
            {"role": "user", "content": "What is 2+2?"},
            {"role": "assistant", "content": "4"},
        ],
    ]
    scores = await ruler(message_lists, debug=True)

[arcticfly]

Instead of assert, could you use an if -> raise block, similar to line 184? A user might otherwise have to do a bit of debugging to figure out what exactly is going wrong.

[vivekkalyan]

Thanks for the suggestion, added raise statements to both conditionals