--n-gpu-layers in llama.cpp: how to set -ngl

What llama.cpp’s –n-gpu-layers (-ngl) puts on the GPU, how to size it from free VRAM and the GGUF file, and what fails when it is set too high.