docs: update Performance section with two-host c7i.xlarge benchmark numbers

[levleontiev]

The Performance section in README.md still shows single-host c7i.2xlarge numbers from March 2026.

Update with new two-host c7i.xlarge × 2 placement group results (eu-central-1):

Latency @ 10,000 RPS:

• p50: decision 304µs / proxy 302µs / nginx 235µs
• p90: decision 543µs / proxy 593µs / nginx 409µs
• p99: decision 2000µs / proxy 1790µs / nginx 1950µs
• p99.9: decision 4000µs / proxy 5120µs / nginx 3620µs

Max throughput: simple 195,000 RPS / complex 195,000 RPS

Also remove token estimation (tiktoken) row — not supported.

Update methodology note: two-host setup, no CPU pinning (Fairvisor and k6 on separate machines).