I heard that the increase in performance from having more than a single GPU are not extremely significant. Having TWO graphics cards in CF/SLI does not necessarily mean double the performance. Although some benchmarks, especially Far Cry 2 show an almost double increase in framerate using Two enthusiast GPU's in CF/SLI.
A single HD 5870 will run Crysis smoothly, Two of them in CF will not double the framerate, but maybe increase it by 10-15 fps in most cases. Unless you're trying to run a game at extremely high resolutions with the AA turned up, Two, Three, or Four GPU's are not essential. Not to mention having to have a beefy and costly power supply.
Some games benefit differently from a multi-gpu setup. Some like FC2 show significant performance gains, while Crysis and the STALKER games only show a minor increase.
FC2 looks graphically equal, if not better than Crysis. Why does is it so much easier to run at high settings than Crysis?
Few questions there, and you seem to have some answers for them yourself.
2 x GPU's in SLI or Crossfire will not double your FPS in almost all cases ( I cant think of any cases, but there might be a few)
If we look at the newest GPU's from nVidia and ATi then we see that nVidia's SLI scales really well, usually 80-90% scaling which means you'd get 80-90% performance increase, ATi Crossfire doesn't scale so well in the majority of cases (going from personal experience and reviews), usually 65-80% scaling.
3 way SLi and Tri-fire (3 way Crossfire) does not scale as well as normal SLi or CF (2 Cards), You could expect roughly a 50% increase (very rough guess) vs. two cards.
Quadfire and Crossfire X ( 4 cards) is usually a complete waste of money as it doesn't really help performance much at all and sometimes even decreases performance vs. 3 cards.
I'd say the sweet spot is either one really powerful card or two good cards that scale well together like the HD5770's or GTX460's.
Some games are coded badly like GTA, Crysis and Metro2033, Farcry 2 isn't so its easier to run on high settings.