[bench] final CUDA benchmark numbers from NVIDIA A10

[peterajhgraham]

Summary

• benchmarks/serving/results.md: adds a new Phase 5 CUDA section recording ~5 ms/batch inference time on NVIDIA A10 24GB (Lambda Cloud), notes the asyncio scheduling bug tracked separately, and confirms the p99 < 30 ms SLO is achievable at this inference time.
• README.md: updates the decoding accuracy table to Cortex-S R² = 0.60 and Wiener filter R² = 0.48 (real trial-aligned A10 numbers); adds a "Speedup (A10)" column to the Triton kernels table showing up to 27× on sparse cross-attention; updates the serving latency note to ~5 ms/batch measured on A10.
• benchmarks/training/results.md: records the final trial-aligned R² of 0.60 (Cortex-S) vs 0.48 (Wiener filter) from the NVIDIA A10 run, replacing the "pending CUDA" placeholder.

Test plan

• Verify README hero metrics table renders correctly (R² values, new speedup column)
• Verify benchmarks/serving/results.md Phase 5 section is accurate and clearly notes the asyncio scheduling caveat
• Verify benchmarks/training/results.md trial-aligned results table reflects A10 numbers

https://claude.ai/code/session_01GVUkAT2SLAqooWAavzTSp7

---

Generated by Claude Code