Path-table walk: readdir-level speed for the count(*) floor #1322

Part of #1304.

SELECT count(*) FROM './*' on a 1M-file flat directory takes 0.61s with #1321 applied. find -maxdepth 1 takes 0.42s, so the budget is 0.46s. Even with no output at all, dirsql misses the 1.1x bar. The only cost native doesn't have is the SQLite insert, so the walk has to come down to native's cost.

How

1. Profile the count(*) path. Break it into readdir, per-entry allocations, sort, judging, row build and SQLite stepping. If what remains after readdir is SQLite cost, report those numbers.
2. Try raw getdents64/openat with reused buffers, and no per-entry PathBuf/String allocation before an entry is known to match.
3. Walk sibling subdirectories in parallel.

What each glob returns, and the order, must not change.