Evaluation: Fix score format

[Prajna1999]

Describe the bug

The evaluation_run table score column stores two different JSON schemas:

• Newer schema: {"traces": [{"scores": [{"name": "...", "value": ...}]}]}
• Older legacy schema: {"cosine_similarity": {"avg": ..., "per_item_scores": [...]}}

This causes frontend rendering issues. Additionally, the older schema only contains cosine similarity scores (missing judge scores),
while the newer schema has both.

To Reproduce

Occurs sporadically. Likely related to the order in which cosine similarity and judge scores are fetched/merged.

Expected behavior

Single consistent schema for the score column across all evaluation runs.

Additional context

Older schema example:

{
    "cosine_similarity": {
        "avg": 0.6414,
        "std": 0.0800,
        "total_pairs": 24,
        "per_item_scores": [
            {
                "trace_id": "9b80f66b-...",
                "cosine_similarity": 0.7379
            }
        ]
    }
}

Newer schema example:

{
    "traces": [
        {
            "scores": [
                {
                    "name": "SNEHA correctness",
                    "value": ...
                }
            ]
        }
    ]
}

Screenshots

[AkhileshNegi]

How the code is set up is when calculating cosine similarity scores, it stores scores in a flat format since cosine similarity is calculated in Kaapi, while the LLM-as-a-judge score is calculated at Langfuse later on, and when fetching scores from Langfuse, it uses a summary_scores array format which overwrite the cosine similarity scores that the cron job had already calculated because the formats were incompatible and the code didn't merge them.