chore: benchmark routing + tool-call validation

[saschabuehrle]

Summary

• unify benchmark model overrides + provider/cost resolution
• enrich benchmark routing metrics + add real-world tool-call benchmark
• fix full benchmark dataset handling and baseline cost calculations
• add mmlu benchmark concurrency and fix bfcl parallel tool-call evaluation

Testing

• python3 -m pytest
• benchmarks: longbench_full, gsm8k_full, mmlu_full, mtbench_full, bfcl_full, ruler_full, truthfulqa, tool_calls(+agentic/realworld), bfcl agentic, basic_usage (py/ts)

[saschabuehrle]

Rebased onto main, resolved all merge conflicts.

Changes during rebase:

• Skipped the 'fix: improve direct routing cost tracking' commit — main now has a more complete pricing_resolver implementation that supersedes it
• Added cascadeflow_latency_ms metric to benchmark results alongside existing diagnostic fields
• Integrated resolve_model_provider/resolve_model_cost utilities with main's benchmark structure
• Combined real-world tool-calling benchmark with main's conditional provider comparison logic

Ready for review.

Greetings, saschabuehrle