Add official RIP implementation

[hanbit-o]

Summary

• publish the core Robust Instant Policy (RIP) Student's-t trajectory aggregation algorithm
• add typed OpenAI Responses API sampling for KAT candidate trajectories
• add a compact banana-picking example that generates one RIP trajectory from 10 demonstrations, with 10 saved GPT-4o candidates for an offline run
• generate 10 KAT candidates by default and include a checked-in successful 55-step banana-lift reference generated with the paper's 40,000-step setting
• add a README-ready 1-by-3 x/y/z-versus-normalized-step plot: KAT points transition from low-weight red to high-weight gray, with the RIP trajectory overlaid in sky blue
• report each KAT's actual fitted Student's-t residual weight, normalized weight percentage, standardized residual, and per-step weights without an extra optimization pass
• pin DINO ViT Features as an external submodule and keep all project-specific integration in a local adapter
• add packaging, citation metadata, tests, and CI

Scope

This is a curated official implementation of the paper's core method, not a full reproduction of all simulation and real-robot experiments.

Validation

• ruff check .
• pytest -q (15 passed)
• offline banana example at 20 optimization steps, including fitted KAT-weight reporting
• Q=10 successful banana reference and README plot generated at 40,000 optimization steps
• uv build
• CFF 1.2 schema validation
• submodule cleanliness and pinned revision checks