feat: Implemented an explicit, configurable context-window override

[ansh-info]

• Summary

Issue - #473

• Add a configurable context_window to RULER (ruler and ruler_score_group), propagating it to LiteLLM as num_ctx/max_input_tokens so Ollama/long-context models aren’t capped at the default 8k.
• Preserve caller overrides: only set num_ctx/max_input_tokens when not already provided via extra_litellm_params.

Motivation

• With Ollama or other long-context backends, LiteLLM defaults to ~8k context unless num_ctx/max_input_tokens are passed. RULER didn’t expose a way to set these, leading to “token count exceeded 8192”
  despite larger model ctx settings.

Details

• ruler now accepts context_window and merges it into LiteLLM params using setdefault.

• ruler_score_group exposes the same knob and forwards it to ruler.

  Testing

• Not run (please run RULER/LiteLLM smoke tests against a long-context Ollama model with context_window set, and verify no 8k cap errors).