CUDA: fix q_nope_absorbed precision for Deepseek 2 Lite f16#13137

JohannesGaessler merged 1 commit intoggml-org:masterggml-org/llama.cpp:masterfrom

JohannesGaessler:ds2l-precJohannesGaessler/llama.cpp:ds2l-precCopy head branch name to clipboard