Fix sRGB & alpha pre/postmult conversion in run.py's PSNR computation (match paper's PSNR numbers)

run.py still contained an early, subtly incorrect PSNR computation that has to do with how prior art in NeRF papers (incorrectly) interprets the PNG alpha channel as sRGB-space rather than linear space. This is further complicated by the fact that NeRF's raymarching procedure also operated in sRGB space (rather than linear), which is both non-physical and introduces a subtle discrepancy with traditional PNG background blending. The new run.py code works around all that -- as the code that generated the paper results already did.

The following line of code will now produce a PSNR of ~36.35, matching the paper and reflecting the actual quality of our results.
`./run.py --mode nerf --scene lego/transforms_train.json --test_transforms lego/transforms_test.json --nerf_compatibility`