actiondfs: shrink cached metadata and simplify inode, parsing, and backing I/O

[dzbarsky]

Summary\n- Deliver 20 separate actiondfs commits reducing cached metadata, protobuf parsing, inode initialization, staged ownership locking, backing I/O, and directory iteration.\n- Shrink cached-child records from 56 to 48 bytes, saving 236,384 bytes across 29,548 LLVM-workload records; decode SHA-256 directly into cached children, avoiding 945,536 bytes of temporary digest copies.\n- Replace production cached-directory atomic exchanges with release stores; the instrumented LLVM run recorded 169,617 directory loads. Allocate staged-directory iteration state only when needed; synthetic VM coverage recorded 13 hits and zero misses.\n\n## Validation\n- Complete supported remote build: 66 targets, both Linux kernels, and every release binary.\n- Complete remote tests: 2/2; standalone timing-parser test: 1/1.\n- Production and instrumented macOS VM end-to-end suites, including sandbox escape, staged filesystem, PID-limit, and timeout recovery regressions.\n- Fresh LLVM VM benchmark: warmup 96.797s / 2,017 executed actions; measured 117.050s / 2,106 executed actions.\n- Parent-main matched CI baseline: Linux actiond/native 215.678s / 217.647s (0.991x); Windows 391.290s / 379.545s (1.031x), with 1,998 executed actions in every comparison. Follow-up PR CI measurements pending.