LLM judge scorer should handle other refusal patterns

[iphan]

Currently the judge scorer only detects "content_filter" refusal from Anthropic. It should be enhanced to support other refusal patterns.
See this comment