I don't know the DirectX terms exactly, but indeed, find an example that does "MRT" (multi render target) so you can draw an object into multiple buffers. I'm pretty sure that nVidia SDK has a lot of examples on that. Once you have set that up, try to render several pixel attributes into your textures. For example:
* texture1: rgb = pixel diffuse color a = specular term
* texture2: rgb = pixel world position a = specular glossiness
* texture3: rgb = pixel normal a = ...?
Note this is just an example. You can compress data such as the normals and positions to get more available for other attributes. But this would be an easy start. Also note that this step does not involve any lighting so far.
Yes. Compare it with photoshop & layers. In the first layer, you have your scenery with its diffuse colors - no lights yet. On a second layer, you draw a red circle that represents a red point light. Set the layer blending mode to "additive" or "light up", or whatever its called in PS. Then on layer 3, you can make another lamp, and so on. Finally, merge all light layers and multiply it with the first lighting. It's not exactly the same, but pretty close.
Blending is not too fast, though additive blending is pretty simple, and I wouldn't worry about the performance unless you want LOT's of lights and/or target lower end hardware. But once you master Deferred Lighting, you could pick up Compute Shaders which allow you you to do all lights in a single pass, applied on smaller tiles on the screen. The Battlefield3 Frostbite engine has a nice paper that explains this "Tiled Deferred lighting". But anyhow, that's for later concern.