feat: Provider-agnostic judge support added to RULER

[ansh-info]

• Summary

Issue: #475

• Make RULER judge provider-agnostic by allowing a LangChain ChatModel to be passed directly (e.g., ChatOllama, ChatNVIDIA) via judge_chat_model, while keeping the existing LiteLLM/OpenAI path intact.
• Maintain context-window passthrough (num_ctx/max_input_tokens) for LiteLLM; the LangChain path is unaffected by LiteLLM defaults.

Motivation

• RULER was tightly coupled to LiteLLM/OpenAI identifiers, blocking clean use of non-OpenAI providers (Ollama, NVIDIA) as judges.

Details

• ruler/ruler_score_group accept judge_chat_model; if provided, use LangChain ainvoke on LangChain messages and parse JSON into Response.
• Keep existing LiteLLM behavior when judge_chat_model is not supplied.
• Context-window overrides still apply on the LiteLLM path; LangChain path bypasses LiteLLM.