[docs] Update README with final CUDA serving benchmark numbers

[peterajhgraham]

$(cat <<'EOF'

Summary

• Replaces MPS placeholder serving metrics with real CUDA A10 measurements from benchmarks/serving/results.md
• Updates throughput (157.3 → 255 req/s), p50 (70.2 → 27 ms), p95 (new), p99 (358.5 → 261 ms), failure count (0/200 → 0/500)
• Adds honest engineering note that the 261 ms p99 is a first-batch initialization spike (CUDA JIT + worker warmup) and steady-state p99 is ~28 ms
• Updates section header from "MPS, concurrency=16" to "NVIDIA A10 (24GB), concurrency=8" to match actual test conditions

Test plan

• Verify numbers match benchmarks/serving/results.md exactly
• Confirm the p99 caveat note is accurate and clearly worded

https://claude.ai/code/session_011ZWNew2AiBDntmujDWfDKm
EOF
)

---

Generated by Claude Code