llama.cpp fixes CUDA argsort corruption bug

llama.cpp has released a patch addressing a memory corruption bug in its CUDA backend caused by improper argument handling in CUB's argsort function. The issue occurred when the DeviceRadixSort::SortPairs API was called with aliased input and output buffers, causing the sort operation to overwrite its own input data mid-pass. This resulted in corrupted permutation indices, particularly affecting top-k operations on large vocabularies (e.g., 248k columns) using CUDA 12.5 and CCCL 2.x. The fix involves using distinct keys-out buffers for all six call sites, ensuring stable inference on NVIDIA GPUs.