Adding score function code (and U-Net code) for pixel single-integrator.

[gchou95]

I more or less copied Yang Song's U-Net architectures verbatim, so I added a compatible variant of the the Langevin dynamics code and created a new NCSN class specifically for images, which has different loss function code.

Let me know if you'd prefer if I try to rewrite everything in the original code structure.

---

This change is 

[hjsuh94]

Thanks Glen! This looks good for our purpose.

I think we might keep the image part of nn.py in a different file? It might make the code a bit cleaner in that file.

Reviewable status: 0 of 8 files reviewed, 1 unresolved discussion (waiting on @gchou95)

---

score_po/nn.py line 489 at r1 (raw file):

    return (sigma_max - sigma_min) * x + sigma_min

############################################ Image-based networks ############################################

Perhaps it's good to separate this part into another file (image_nn.py)?