[PB] Clean up grade function flow

[thesofakillers]

goal was to get paperbench.nano.task.PBTask:grade() as clean and minimalist as possible, ideally so that you can read it in natural language and know what's going on.

• Various variable renames and docstring/comment updates
• Removed a misleading comment
• Abstracting pieces of logic to more readable helper functions
• Keeping function-specific logs inside the functions themselves
• Some _run_reproduce-specific cleanup:
  • we were re-selecting the checkpoint and running checks on its existence in run_reproduce, despite JUST doing that in grade, so we removed it since it's redundant
  • Then we mimic the shape of how we call judge, i.e.
    • starting the computer inside the fn, so we only have it running for when it is needed
    • use a _should_reproduce helper function to handle the conditionals