Use real MS-MARCO queries in benchmarks

[tjgreen42]

Summary

Our MS-MARCO benchmarks currently use hardcoded query strings like "coffee" and "how to cook" instead of the actual MS-MARCO query set. We should use a subset of the real queries for more representative benchmarks.