Statistics for multiple reruns

[ageneric]

Additions:

• aggregate_regrading.py to perform repeated runs and collect statistics

Changes:

• Updated documentation
• Relocated unused configs to legacy/

Technical changes:

• For readability, renamed "token_usage" to "text_generation" to better describe the functionality the script provides
• Removed duplicate 'local' copies of functions related to reasoning models
• Instead of setting reasoning effort to "medium", simply use the default reasoning effort (which is "medium" anyway)

• Will upload result files in next commit (didn't want to bury the code changes in all the added files)