Bump llama.cpp to b1842 and add new cuda lib dep

Upstream llama.cpp has added a new dependency with the
NVIDIA CUDA Driver Libraries (libcuda.so) which is part of the
driver distribution, not the general cuda libraries, and is not
available as an archive, so we can not statically link it.  This may
introduce some additional compatibility challenges which we'll
need to keep an eye on.

llm/generate/gen_linux.sh

@@ -141,6 +141,7 @@ if [ -d "${CUDA_LIB_DIR}" ]; then
         ${CUDA_LIB_DIR}/libcublasLt_static.a \
         ${CUDA_LIB_DIR}/libcudadevrt.a \
         ${CUDA_LIB_DIR}/libculibos.a \
+        -lcuda \
         -lrt -lpthread -ldl -lstdc++ -lm
 fi