fix: More conservative VRAM estimate for base models with weights

Calculate: base_vram + 2GB + 50%
This ensures a 14B model estimated at 18GB will require ~29GB instead of 22.5GB.