vLLM#56535: FlashInfer b12x MoE runs W4A16 checkpoints as W4A4

FlashInfer flashinfer_b12x MoE backend runs W4A16 checkpoints as W4A4. Root cause: vLLM knows checkpoint is W4A16 but never tells FlashInfer. FlashInfer defaults to fp4 activation precision. Fix: pass quant_mode and activation_precision. GSM8K: 0.9037 (fix) vs 0.9136 (marlin). Demonstrates quantization accuracy loss from kernel misdispatch.