To answer your first question: only 3dREMLfit deals with temporal autocorrelation.
To answer your more philosophical question is more complicated.
In the past, 3dREMLfit was much slower than 3dDeconvolve, and computers were also much slower. The result was that there was a significant time penalty for using 3dREMLfit. And at the **group level**, there is very little difference between the group analysis results using 3dREMLfit and using 3dDeconvolve for the individual subject task analyses.
In the present, I have speeded up 3dREMLfit, and also computers are much faster. The upshot is that I personally always use 3dREMLfit for my task analyses. Also, 3dREMLfit has a capability that 3dDeconvolve does not have: voxel-wise regressors, which means that ANATICOR can be used with 3dREMLfit.
