voice-bench — per-GPU configs

This directory mirrors packages/inference/configs/gpu/ for the benchmark harness. The four matrix entries in matrix.json map a voice-bench invocation to a (GPU, bundle, ctx_size) triple.

The actual llama-server flag set comes from packages/inference/configs/gpu/<gpu>.json via the gpu-autotune.ts helper in @elizaos/app-core. We do not duplicate that data here — only the bench-runner-specific knobs (turns, warmup, fixtures) live here.