Benchmark harness on /tmp hides fsync cost (tmpfs)

A storage benchmark harness that puts its scratch export under /tmp silently measures with free fsyncs on systems where /tmp is tmpfs. Durability-heavy designs (write-ahead logs, group commit) then look like pure overhead, and tuning against those numbers optimizes the wrong thing. Make the scratch dir configurable, default it to the working directory, and keep it on a real disk (check with findmnt -T <dir>). Designs can also adapt at runtime: time the first few fsyncs, and skip the log when they're about 50 µs or less.