Update README for AI4Law @ ICML 2026 (new title, citation-faithfulness, Figure 1)

[boqiny]

Updates the README to match the accepted AI4Law @ ICML 2026 camera-ready version of the paper.

What changed

• Title / framing: retitled to Legal Multi-Agent System with Agentic Search and Citation-Faithfulness Audit and added an ICML 2026 AI4Law acceptance note.
• New section, Citation-faithfulness audit: documents the VerifyAgent, the six-class claim taxonomy, strict-ALCE scoring, cross-provider judging, and Faith-Search. Also states the narrow scope of "faithfulness" (reachability plus snippet entailment, not doctrinal good-law validity).
• Evaluation summary: now reflects the paper's headline finding. Accuracy stays flat while the multi-turn loop lifts strict citation F1 from 0.13 to 0.25 and cuts the no-citation rate from 34% to 13%; Faith-Search drops the unreachable rate below 1%.
• Figure: replaced the workflow image with Figure 1 from the paper (assets/workflow.png), showing Simple Mode and Multi-Turn Mode with the Judge Agent and optional Faith-Search.
• Citation: updated to an AI4Law workshop entry.

Please verify before merging

• Citation author order / venue string: used Boqin Yuan and Ziqi Wang (OpenReview order) and booktitle ICML 2026 Workshop on AI for Law (AI4Law). Confirm the exact workshop name and whether to also keep the arXiv @misc entry.
• arXiv title: the linked arXiv record (2509.00761) may still carry the previous title; update it if you want the link and README to match exactly.
• Benchmark script paths (run/single_turn_pipeline.py, eval/run_eval.py) were left unchanged from the previous README and not re-verified against the current tree.