actiondfs: reduce lookup, allocation, locking, and backing-file overhead

[dzbarsky]

Summary

• Deliver 31 independently reviewable actiondfs improvements across inode lookup, digest/cache records, backing-file I/O, staged files, locking, allocations, and memory use.
• Store SHA-256 digests in binary; reuse spare cached-child storage for filenames; reject full digest-key mismatches before comparisons; eliminate per-node mutexes, backing-file references, workqueues, parser stack frames, and unnecessary owner/idmap conversions.
• Keep credential contexts off synchronous read/write hot paths, simplify generic attributes and iterator clamping, and avoid redundant privilege processing without changing setuid/setgid semantics.
• Preserve hashed inode identity and filesystem watches; add staged zero-length/nonempty write permission regressions; refresh synthetic stress and canonical LLVM timing reports.

Validation

• Supported remote repository build: 66 targets; remote repository tests: 2/2.
• Production and instrumented Virtualization.framework VM suites, including sandbox escape, setuid/setgid write/copy, stress, PID, and timeout regressions.
• Standalone stress timing-parser test.
• Fresh LLVM VM smoke: 2,017 warmup actions in 92.411s; 2,106 measured actions in 109.140s. Before the latest ten commits the same action counts measured 97.375s/144.161s; local host contention varies.
• Matched Linux CI: actiond/native 321.388s/308.108s, 1.043x; main 281.169s/279.225s, 1.007x; 1,998 executed actions each. Historical Linux comparisons across runners range from 0.985x to 1.044x.
• Matched Windows CI: actiond/native 252.985s/260.467s, 0.971x; main 327.303s/332.624s, 0.984x; 1,998 executed actions each.
• All five GitHub Actions checks passed. The first Linux runner reproduced an existing main gRPC keepalive failure; its targeted fresh-runner retry passed.