add some features for torch training

[baibaidj]

1. enable amp to accelerate training and reduce GPU memory footprint.
2. add instance norm to convblock for better performance
3. add two parameters to argparse, --nb_conv_per_level for better control of the model size, and --mean_std for normalizing data on the fly.
4. add data flip in data generator (scan_to_scan func).
5. adjust parameters for nn.Upsample and nnf.interpolate to remove warnings