Assuming you're correct about the 1.5--4.5 range (I don't follow the IPCC that closely), then unless there is some reason to expect an asymmetric uncertainty distribution, then 3 C would still be the estimate--it's the average of 1.5 and 4.5. They are choosing to emphasize how large the spread might be, rather than the most likely value.
The uncertainty range comes from running multiple models multiple times and each time you run the model with slightly different initial conditions, you get a different result. Numerical weather prediction models have the same problems when they're forecasting out past a week or so, which is why ensemble forecasting is useful at that time range. What the climate models really need is to be run globally at cloud resolving resolution, but that will require advances in computing power.