NVFP4 vs FP8 vs GGUF on a 16 GB Ada GPU: NVFP4 decodes 2.3x faster and ties BF16 on GSM8K