Add support for the new PAE JSON format

[Augustin-Zidek]

• Adds support for the new PAE JSON format while also keeping backwards compatibility with the legacy format.
• The new format makes the PAE JSON about 4x smaller when stored compressed, even smaller when stored uncompressed (we don't store redundant residue indices and the predicted aligned error is rounded to integers).
• The new format parses about 3x faster.
• I kept the dtype of the matrix as np.float64 to not break any existing code, but since PAE is now integers, np.int32, np.float32 or even np.float16 could be used is RAM usage is too high.

[tristanic]

Thanks!

Will merge this... but can you tell me a little more about this?

and the predicted aligned error is rounded to integers

As written, that's a bit worrisome for a few different reasons. First is the 5-fold loss of precision (previously I believe it was reported in steps of 0.2 Angstroms, now just 1 A steps). Second, PAE values less than 0.5 would presumably round down to zero, causing potential divide-by-zero issues for any code that wants to use 1/(PAE). You could get the same level of compression without loss of precision or introduction of zeros by storing the data as int(round(pae*10))... is that a possibility?