some improvements and questions about psnr

[lunasdejavu]

I think the resize input_transform should be bicubic since the default value is bilinear,
and centercrop should be changed to RandomCrop()
and some of the calculation of the PSNR in test should be normalize first and use MSELoss() not L1Loss() directly. and there still might be some problems that I can't achieve PSNR >30 by this project