Unsloth achieves 2.5x faster Qwen3.6 NVFP4 quants

Unsloth's NVFP4 quantized Qwen3.6 models achieve 2.5x speedup over NVIDIA's version with W4A4 tensor cores, with 1.56-1.79x gains on 35B-A3B. FP8 KV cache calibration is included and no accuracy loss is claimed.