Improving the accuracy of single-trial fMRI response estimates using GLMsingle