nx_f32_llm_bench_test.nx

nx_f32_llm_bench_test.nx -- parked LIVE-FIRE perf harness for real Qwen2.5. Reads /tmp/nx_real_model.gguf, loads dims + weights + vocab, runs nx_f32_llm_run_v2 with a fixed prompt + 32 new tokens, captures wall-clock via sys_now_ms, and writes a TSV row to /tmp/nx_f32_llm_bench.tsv with: prompt_n / max_new / n_emitted / elapsed_ms / tokens_per_sec_milli (= tokens * 1000000 / elapsed_us). Substrate-honest scope: * This is the HARNESS, not a perf result. No claim is made about Nishi tokens/sec until this smoke runs PASS on the real file and produces measured numbers. * No comparison to llama.cpp is made here. A paired bench against llama.cpp on the same Q4_K bytes on the same hardware would be a separate brick. Per feedback-no-strawman-perf-comparisons: NO inflated claims. Smoke wrapper skips gracefully when the file is absent.