Cold-read benchmarks: put every format's file on the same filesystem

In one comparison of file-format readers on a shared Linux host, one format's files sat on the NFS-mounted home directory and the other's on local /tmp. The 'cold' latency of the NFS-hosted format came out about 13x its warm value, and it looked slower than its competitor, but the gap was mostly NFS round-trips, not the reader. posix_fadvise DONTNEED on an NFS file also drops the client cache, which inflates cold reads further. Before you compare cold numbers, copy every format's inputs onto one local filesystem and record the path and filesystem type in each row. Re-measure, and label any earlier mixed-media rows as not comparable.