Add cost and token tracking per benchmark run

[rajkumar42]

Summary

Add instrumentation to track token usage (input/output) and estimated cost per benchmark task and per full run. This enables cost comparison across providers and helps users estimate expenses before running large benchmarks.

What needs to happen

• Capture input and output token counts from each LLM call
• Map token counts to estimated cost using per-provider pricing
• Aggregate totals per task and per full benchmark run
• Output a cost summary at the end of each run (total tokens, total cost, avg cost/task)
• Save cost data alongside benchmark results for later analysis

Example output

Benchmark complete: 162 tasks
Total tokens: 1,245,000 (input: 980,000 / output: 265,000)
Estimated cost: $4.82
Avg cost/task: $0.03
Provider: claude-sonnet-4-6

Acceptance criteria

• Token counts are captured per LLM call
• Cost summary is printed at end of run
• Cost data is saved to results file