Ideally, you would use -stim_times_AM3 -- but that option doesn't exist at the current time coordinate. If you use -stim_times_AM2 twice, as you describe, and then use -GOFORIT to push on past the collinearity warning, what will happen is that the two identical regressors will each get half the beta weight due to the mean response. And they will each have a small t-statistic. If you don't care about the mean response at all (e.g., run that separately to get an activation map for the main effect), then you can proceed. But you should be careful to examine your results to make sure that they make sense.

An alternative would be to run with -stim_times_IM and get the coefficients for each event separately. Then you'd have the problem of regressing the collection of event amplitudes in each voxel against your A and B parameters. That itself could be done with 3dDeconvolve, using the appropriate pieces of the -cbucket output from the first run as the input to this second run, and using -stim_file inputs with the A and B amplitudes, and using -polort 0 to pick up the mean.

Other clever things could be done with the -stim_times_IM output -- that's up to your imagination.