Update predict_structure.py

[DimaMolod]

fast resume

[jkosinski]

It looks good, but it would still be slow on the first restart because, if I understand correctly, ranking_debug_temp.json will still need to be created by iterating and reading all PKL files, no? In my experience, I rarely restart more than once. I think then ranking_debug_temp.json needs to be created and updated in the predict() function.
An old implementation of this, which I had implemented and later removed (because I thought reading pkl was a better idea), was in the old version of the script:

AlphaPulldown/alphapulldown/predict_structure.py

Line 155 in 60807c4

with open(temp_ranking_output_path, "w") as f:

Note also that the ranking_debug_temp.json file is removed after the job completion.

But note also predict() is the part that Valentin is also changing now.