Normal maps are almost always more efficient than polygons..plus they are far more flexible than just adding polygons.. the quality is so much better with per pixel lighting and normal mapping, that the question is moot. Normal maps give you ability to add a huge amount of detail.

Normal mapping/per pixel light shaders are run on the GPU, not the CPU. The reason they are more efficient than adding far more polygons is simply that running a pixel shader is faster than transforming and drawing thousands of polys. If you have 5 characters on the screen of 3000 polys each, thats about 15,000 polys, minus backfaces, and LOD. Now if those models were 10,000 polys ach, then you'd have 50,000 polys. If you can use normal maps to make the 3000 poly model look essentially the same as the 10,000 poly model, the difference are huge.

Per-pixel lighting/normal mapping shaders won't work with the static lights. you will need to use dynamic lights, as the engine automatically passes the position, color, and range of each dynamic light to the shader.