docs(bench): refresh inference performance matrix

[bsbodden]

Summary

• refresh the controlled Qwen3, SmolLM2, and MiniCPM5 cross-engine matrix
• separate measured Java progress from Ollama run variance
• record current prefill, TTFT, decode, memory, and output-parity results

Validation

• ./gradlew check
• git diff --check
• clean controlled VPS run: 8 vCPU AMD EPYC Milan, two warmups, ten trials, identical GGUF bytes