Benchmarking Qwen 3.8 27B on RTX 5090 and beyond — VRAM capacity alone can't overcome severe software and inference engine bottlenecks