perf(tfork): scope pre-fork writeback to the source filesystem

[yiying-zhang]

Summary

• replace host-wide sync with syncfs on the source container rootfs;
• preserve a compatibility global mode and an experimental none mode;
• make the selected durability boundary explicit through
  PODMAN_TFORK_SYNC_MODE=syncfs|global|none;
• add dirty buffered-write, shared-mmap, rename, sparse-file, and isolation
  coverage.

Why

The old global barrier runs while the source is frozen and can flush unrelated
filesystems and workloads. syncfs preserves filesystem-wide writeback for the
source backing filesystem without paying for every filesystem on the host.

Performance

Ten interleaved real-Webtop pairs measured:

Mode	Wall median	Mean
host-wide sync	1.535 s	1.557 s
source syncfs	0.850 s	0.848 s
saving	0.685 s	0.709 s

CRIU latency was unchanged, confirming that this removes outer filesystem
writeback rather than snapshot work.

Correctness and durability

All 30 synthetic forks passed across global, syncfs, and no-sync modes. The
matrix covered 64 MiB buffered dirties, MAP_SHARED dirties without msync,
rename, a 1-GiB sparse file, inheritance, and child/source isolation.

syncfs still flushes the complete filesystem, not only the rootfs directory.
Sources using multiple independent mounts require one barrier per participating
filesystem. The no-sync mode is not a durable production default. A disposable
VM power-cut test remains the final crash-durability gate.