v1.7.2 - Fix CPU-only model load failure

Fixes 'Failed to load model' on the CPU-only backend. main_gpu/split_mode are now only applied when GPU offload (n_gpu_layers > 0) is requested, since llama.cpp rejected main_gpu=0 when 0 GPU devices are available. Adds native llama.cpp log callbacks and surfaces load failures with the log path. Requires .NET 10 Desktop Runtime (win-x64).