Add benchmarks: prompt size with vs without SkillMesh

[varunreddy]

Summary

Create a benchmark script that measures prompt token counts, retrieval accuracy, and cost savings when using SkillMesh routing vs loading all tools into the prompt.

Acceptance Criteria

• Create a benchmark script in benchmarks/ or scripts/
• Measure token counts for full catalog vs top-K routed context (K=3, 5, 10)
• Report accuracy: does the top-K include the correct expert for a set of test queries?
• Include at least 20 test queries across data, ML, DevOps, web, and cloud domains
• Output a summary table suitable for inclusion in the README
• Document how to run the benchmarks