Llama.cpp doubles prompt speed via PCIe optimization

A simple environment variable change fixes a common bottleneck for consumer-grade multi-GPU setups where secondary slots are often electrically limited. Prefill performance is highly bandwidth-sensitive, making PCIe lane distribution critical for large models. Reordering GPUs ensures the primary device for initial processing sits in the x16 slot rather than an x4 chipset-linked slot, which is particularly beneficial for Mixture-of-Experts models that require frequent data shuffling. Users can verify bandwidth caps via nvtop or lspci to unlock this "free" performance on standard desktop platforms like the X570.