In fact, the input gamma of 2.2 will "un-gamma" your image before they are sent to the render engine, that way it "sees" the color in it's own color space (render engine color space = gamma 1, origin of the name "linear"). 

NB: of course if you were to hand paint or generate procedural texture and have them looking good on a display set to a gamma of let's say 1.8, then you would need to set the input for that particular image to 1.8 too.

As for the output gamma, it really only depends on what you want to do with the picture you render.

1. you want to work on it in a program that can handle 16 or 32 bits images: in that case, the best is to save it as a float image with a gamma of 1. (Actually here, the gamma isn't very important, it will only tell your software how to show the picture but all data are conserved anyway).

2. you want the picture to look at its best on your computer and don't really plan on posting it anywhere: choose to save for the gamma of your monitor.

3. you plan on posting the picture on the web: the safest choice here is probably a gamma of 2.2 since it's the most commonly found (at least among people who actually calibrate their monitor :) )

In my opinion, the best choice is 1 (unless you really don't plan on doing any kind of post-work or compositing on your image. This option allow you to have the full range of data produced by the render engine at your disposal and, for example change the exposition of your image drastically without problems.

I don't know who wrote the article for 3D world, but I can assure you he/she was wrong concerning the input gamma. You want to "feed" your render engine with what it expect to find :)