GPU Inference RTX 5090 + Nemotron Nano 9B v2 Japanese on vLLM 0.15.1 — Real Benchmarks, Reasoning Parser Fix, and Why I Skipped TRT-LLM Loading...