Benchmark and improve Nix NAR unpacking

[siraben]

Summary

• Add a repeatable Nix NAR unpack benchmark that prepares a riscv64 fakefs with Nix installed and a baked local NAR fixture.
• Speed up fakefs metadata-heavy unpacking by tracking occupied stat-cache slots and updating matching cached inode entries instead of scanning/clearing the whole cache.
• Batch fakefs metadata writes for exclusive file creates, mkdirs, symlinks, and fd metadata changes; pending paths/stats rows are overlaid for reads and flushed to SQLite in bulk.
• Reuse cached fakefs fd metadata during fsetattr, and lower the direct-I/O threshold used by the measured candidate.

Benchmark

Fixture: RESTORES_PER_RUN=5 RUNS=5 WARMUPS=1 python3 benchmarks/nix_unpack/run.py

baseline_x5             avg 3.683045s
candidate_used_slot_x5  avg 3.635887s
path_stat_writeback_x5  avg 3.000889s

The post-run persistence smoke test confirmed 3256 restored paths and 3256 matching stat rows were flushed to meta.db.

Rejected experiments are documented in benchmarks/nix_unpack/README.md.

Validation

• python3 -m py_compile benchmarks/nix_unpack/prepare.py benchmarks/nix_unpack/run.py
• ninja -C build-perf-rv64 ish
• meson test -C build-bench-rv64 rvinterp-selftest rvdecode-selftest
• Persistence smoke test: one full restore, then sqlite3 meta.db counts for restored paths/stats both returned 3256