(feat)Rich logging

[e-dobrowolska]

This PR brings an optional, richer console logging experience to all the benchmark inference scripts.

What changed

• Added a new shared logging module: benchmarks/utils/console_logging.py
• Conversation output is now much richer and shows tool/message/error events live in the console (via benchmarks/utils/conversation.py)
• Switched how child process logging is set up: now uses the new centralized helper in benchmarks/utils/evaluation.py
• Added end-of-instance summary lines (summarize_instance(...)) in all the main benchmark runner scripts:
  • benchmarks/swebench/run_infer.py
  • benchmarks/swebenchmultimodal/run_infer.py
  • benchmarks/swtbench/run_infer.py
  • benchmarks/multiswebench/run_infer.py
  • benchmarks/commit0/run_infer.py
  • benchmarks/gaia/run_infer.py
  • benchmarks/openagentsafety/run_infer.py
• Added some docs: see benchmarks/utils/logging.md for how it looks and acts

Behavior impact

• By default, everything works as before.
• When enabled, you'll see:
  • a startup line for each instance
  • live updates for tools, messages, errors, etc.
  • a short end-of-instance summary
• File logs are still written as before:
  logs/instance_<id>.log
logs/instance_<id>.output.log

How to enable

• export RICH_LOGGING=1

[juanmichelini]

tested both with and without RICH_LOGGING=1, LGTM

[juanmichelini]

thanks @e-dobrowolska !