Add PAPER_v2.md: adaptive routing negative result paper

Add v2 ablation results: adaptive routing scores 40% EM (vs 46% v1)

Implement RLM-Memory v2: query-type adaptive retrieval

Rewrite PAPER.md to match LaTeX paper with all verified results

Add RAG baseline evaluation and update paper/README

feat: add RAG baseline (text-embedding-3-small + gpt-4o-mini)

fix: correct Hindsight model name (not Gemini-3, open-source 20B+)

docs: update README with verified 4s parallel latency (54x speedup)

docs: update paper with verified 54x parallel latency results

perf: verify 54x latency reduction with parallel sub-agents

feat: parallel sub-agent execution via ThreadPoolExecutor

Add files via upload

Initial release: RLM-Memory — scalable conversational memory via sub-agent delegation