Add sync_fs() to service FUSE_SYNCFS

[edmc-ss]

Summary

Adds support for servicing FUSE_SYNCFS — the opcode the kernel issues for syncfs(2)/sync(2) — plus the opt-in FUSE_HAS_SYNCFS INIT capability that lets a privileged /dev/fuse server receive it.

• New Filesystem::sync_fs(&self, req, ino, ReplyEmpty) trait method; defaults to ENOSYS, which the kernel treats as a permanent success (same convention as fsync/statfs).
• FUSE_SYNCFS = 50 opcode, a header-only SyncFs request, its Operation variant, and request parsing/dispatch/Display.
• InitFlags::FUSE_HAS_SYNCFS = 1 << 43 (protocol 7.46), requested by a server via KernelConfig::add_capabilities().

Dependency on the Linux VFS fuse driver (please read)

This is the important caveat. Although the FUSE_SYNCFS opcode has existed for a while (added for virtiofs), the in-kernel fuse driver only propagates syncfs()/sync() to the userspace server (fc->sync_fs) for virtiofs and fuseblk mounts. A plain /dev/fuse server never receives it today — regardless of what it advertises — because an untrusted server could stall sync() indefinitely.

Consequences:

• On current released kernels, this PR immediately benefits fuseblk fuser servers (they already receive FUSE_SYNCFS), and is a harmless no-op for plain servers.
• For plain /dev/fuse servers to receive it, the kernel needs the in-flight opt-in: a new FUSE_HAS_SYNCFS INIT flag, honored only when the server opened /dev/fuse with CAP_SYS_ADMIN in the initial user namespace (the same privilege that mounting virtiofs/fuseblk already requires). That series — "fuse: allow FUSE_SYNCFS for privileged userspace servers" — is on LKML (patch, LWN summary) and not yet merged. The 1 << 43 bit and protocol minor 7.46 here match that patch's UAPI.

Because KernelConfig::add_capabilities(InitFlags::FUSE_HAS_SYNCFS) returns Err when the running kernel doesn't offer the flag, a server can cleanly detect an unsupporting kernel and fall back (e.g. to an out-of-band flush) rather than silently never being called.

Design notes

• FUSE_KERNEL_MINOR_VERSION is intentionally left at 40. The kernel gates this flag on the opener's privilege rather than the negotiated minor, and it rides in the high-32 flags2 (already carried via FUSE_INIT_EXT), so it works without claiming the intervening 7.41–7.45 features fuser doesn't yet implement. Happy to bump the minor if you'd prefer.
• SyncFs is modeled on StatFs: filesystem-wide and header-only (the wire fuse_syncfs_in is a single reserved u64 padding word, so there's no argument to expose).
• Mirrors libfuse's low-level syncfs op, which likewise replies ENOSYS when unimplemented.

Testing

cargo test (including a new FUSE_SYNCFS parse test), cargo clippy --all-targets, cargo fmt --check, and cargo doc (with -D rustdoc::broken_intra_doc_links) all pass on Linux (aarch64, toolchain 1.85). Not exercised against a live mount, since the privileged-plain-server kernel path isn't in a shipping kernel yet (see above).