[PB] Cleanup reproduce flow

[thesofakillers]

Our reproduction execution logic was clunky and unwieldy to work with. This PR streamlines it with a bundle of quality-of-life improvements:

• Add a warning when salvaging will be skipped due to timeout configurations (otherwise it would silently skip)
• Improve logging throughout the code
• Remove stale TODOs
• Ensure explicit support for blobfile
• Consolidate ReproductionMeta and ReproductionOutput into a single structure
• Pass dataclass instances directly instead of converting to and from dictionaries
• Simplify reproduce_on_computer by removing:
  • Installation steps that can be handled in the Dockerfile
  • Inline try/except blocks (this logic can be moved elsewhere)
• Defer computer start until it’s actually needed
• Move run_reproduce_on_computer to paperbench/scripts/run_reproduce.py so all core reproduce logic is in one place
• Address an existing TODO and add a new TODO to integrate with the ComputerInterface API