Retrying relevance score prompt if score was a float or response isn't valid JSON

[jamesbraza]

Our LLM response --> Context logic is a bit fragile:

• Pydantic error when parsing context #747 shows the LLM responding with a float relevance score.
• Parsing JSON error when increasing evidence_k with mixtral:8x7b #736 shows the LLM failing to respond with valid JSON (though actually the issue seems to have been a token limitation).

What currently happens is this blows up Docs.aquery.

However, this should not be a critical failure because:

1. It's not in the caller's (e.g. an agent) control, it's a failure internal to paper-qa.
2. We can retry the summary prompt once or twice, saying something like:
   • "We failed to parse valid JSON according to exception {exc!r}. Please respond with JSON."
   • "We failed to extract a valid score according to exception {exc!r}. Please provide an integer score in 1-10."