Learning Spatially Varying Pixel Exposures for Motion Deblurring