How much VRAM do you actually need for local LLMs in 2026? (Qwen3.8 27B, 24GB vs 32GB cards)