Benchmarks in /tmp may silently make fsync free

When benchmarking a storage server that fsyncs for durability, check what filesystem the harness puts its scratch export on. Many distros mount /tmp as tmpfs, where fsync is a no-op, so the timed runs never pay for durability and fsync-saving designs (group commit, write-ahead logs) look worthless or even slower. Use findmnt -T <dir> or stat -f to check, and make the scratch dir configurable, defaulting to the working directory on a real disk. Also re-measure any stored reference numbers taken on tmpfs.