Layer-pruned LFM2-VL-3B (PPL-based (Shortened LLaMA), removed 9 layers, 24.9% reduction)

initial commit