perf(walk): improvements to the walker hot path

[arcuru]

More changes as a stack on top of my previous #574. Ideally this would be stacked commits but right now this will show the prev commit inside this PR...

Happy to split/rebase as requested, though I am also fine if you pull/modify yourself.

Five small reductions on the per-entry walker hot path. None of them change observable behavior they just remove redundant work. They compound for noticeable wins on tree-heavy and wide-directory walks.

1. Per-dir stat cache. walk_dir() was statting each directory for its is_dir() check, then build_node() statted it again to populate the Node. Cache the parsed metadata tuple in PendingDir via OnceLock. One statx saved per directory.

2. Cache entry.path() / entry.file_type(). Both were called 2-3× per entry, and entry.path() allocates a fresh PathBuf each call. Compute once in process_entry and thread through. Also gate is_ignored_path on a cheap empty-check so walks without --ignore-directory skip the HashSet probe entirely.

3. Skip ignore_file on default walks. Precompute has_any_filter when constructing WalkData. In the hot loop, three-way branch: full ignore_file when filters are active, inlined dot-file check when only --ignore-hidden is on, otherwise skip. When filters are active, also coalesce the duplicate get_metadata calls inside ignore_file, replace path.is_file() stats with the cached file_type, hoist fs::canonicalize out of the ignore_directories loop, and thread the prefetched tuple into build_node so each entry is statted at most once on filter-active walks.

4. FxHash for inode dedup. clean_inodes hashes 25M (inode, dev) pairs on my test /nix/store. SipHash's DoS resistance is pointless for primitive keys from our own syscalls. Adds rustc-hash = "2", swaps HashSet for FxHashSet. This can be inlined to avoid the dep by defining a simple hasher.

5. Skip per-file statx in --filecount mode. Under -f, every file's size is overwritten with 1, so the per-file statx is removable. The only field the dedup still needs is (inode, dev): inode comes from DirEntry::ino() (filled by getdents64), dev from the parent directory's cached tuple. Gated to -f without -L and without metadata-needing filters, so output is byte-identical to the previous stat path. Unix-only; Windows DirEntry has no cheap d_ino analogue.

Benchmarks

Cumulative for this batch of changes. "Before" is inclusive of #574.

host: 24 CPU; hyperfine --warmup 2 --runs 10 for synthetics, --warmup 1 --runs 3-5 for large targets

target	before	after	speedup
balanced (1000 dirs / 20k files)	9.3 ms	8.8 ms	1.06×
wide_flat (1 dir, 100k files)	76.9 ms	72.9 ms	1.06×
deep_narrow (1500-deep chain)	139.3 ms	98.9 ms	1.41×
-e "\.rs$" regex filter on dust src	35.7 ms	16.9 ms	2.11×
-x dust src	48.3 ms	30.4 ms	1.59×
-f balanced	9.6 ms	7.7 ms	1.25×
-f wide_flat	82.4 ms	74.8 ms	1.10×
~ (273k entries)	182 ms	173 ms	1.06×
/nix/store (7.5M+ entries, ZFS)	33.5 s	25.6 s	1.31×
-f /nix/store	25.6 s	18.8 s	1.36×

Default-walk synthetics (balanced, wide_flat, ~) are tighter on this run. Those workloads are already close to the syscall floor; the per-entry CPU savings show up in user time more than wall time. I also sampled the system times for these runs, and they are more dramatic where they apply: -e -75% sys, -x -41% sys, -f /nix/store -61% sys. strace confirms -f wide_flat goes from ~100k statx calls to ~10.

Further Work

One more potential perf change stacked on top of this:

• AT_STATX_DONT_SYNC on Linux. Biggest win on network filesystems (~3.8× on an 11 TB NFS mount in my testing); helps cold-cache local in smaller amounts. dust's default std::fs::Metadata uses AT_STATX_SYNC_AS_STAT, which for a read-only walker is stricter coherence than we need.

That is a small behavior change though, which may or may not be acceptable.