add quantization_config.ignore=1 (downstream audit fix)

Replace INT4 vision with BF16 originals (config.json)

Gemma 4 26B MoE AWQ: forced-routing GPTQ calibration for all 128 experts, 30 tok/s on RDNA4