With the stimuli like yours, you probably have two approaches:
(1) As you indicated, use the '-stim_nptr k 5' for all the stimuli by specifying those 5N timepoints (assuming you have N TR's in the input dataset) . And you would have to set up the '-stim_minlag k #' and '-stim_maxlag k #' options appropriately since now the time steps are in multiples of TR/0.5. Check the manual for 3dDeconvolve for details about this.
Apart from processing speed, there is some other difference on the various divisions of TR. For example, the finer you divide the TR, the more parameters you would have to estimate, and thus more loss of the degrees of freedom during deconvolution/regression analysis.
(2) If you are not really interested in modeling impulse response function for each stimulus, you may use the '-tstim' option in waver with stimulus files (in real time of those events, not in 0's and 1's) with some assumption about the hemodynamic response (such as Gamma, tent, MGH, etc.), and then run 3dDeconvolve with the regressors generated from waver.
Gang