Measured A/B on the self-hosted setup: warm-bench restore 85s vs full
bench init 82s — no gain (slightly slower). bench init is already fast
because the uv/pip caches are mounted warm, so the cache only replaces a
~13s init while adding a ~200MB untar and still running bench build.
Drop BENCH_CACHE_DIR so warm-bench stays inert (the helper functions
remain, matching develop's install.sh).