This protects against deadlocks if the kernel calls back into
userspace in the notify handler.

To update, change

   nodefs.MountFileSystem(dir, fs, opts)

to

   nodefs.MountRoot(dir, fs.Root(), opts)

This makes it impossible to forget to create an Inode to match the
FsNode.

This is an API change. To update, update function signatures, and change
  
  me.Inode().NewChild("name", isdir, child)
  return child, fuse.OK

to

  return me.Inode().NewChild("name", isdir, child), fuse.OK

in Mkdir, Mknod, Create, Symlink and Link methods.

This obviates the nodefs.FileSystem type. To adjust for this API change,

  * call nodefs.MountRoot() on the Root() node of a to-be-mounted
    filesystem.

  * Move OnMount/OnUnmount methods to the type of the root node.

  * Stop embedding NewDefaultFileSystem() into file system
    objects. Usually FS objects no longer have to be exposed.

NewChild also adds inode as a child.

Code that did:

  ch := n.Inode().New(isDir, fsNode)
  n.Inode().AddChild(name, ch)

should be changed to look like

  ch := n.Inode().NewChild(name, isDir, fsNode)

If we wait longer, we will probably wait indefinitely.

Versions have to integrate with the build/version control system,
which the go tool does not provide.

This could happen to unionfs with ClientInodes set, and manipulations
that skipped the backing store.  Added a test.