* In-process mounts are now always on synthetic inodes.  FileSystems
  no longer have to generate GetAttr() responses for mount points.

  Advantages:

  - Simplifies FileSystems: there is no need to represent the state
    between mountpoint appearing and Mount() succeeding, and mount
    points appear automatically in ReadDir().

  - Mount point inodes are always associated with one filesystem over
    their lifetime.  This simplifies synchronization around unmount calls,
    which could previously lead to "unlock of unlocked mutex" panics.

* (Un)mount always takes 2 locks: treeLock from the parent, and the
  treeLock of the filesystem itself, as the parent/child relation
  crosses the boundary between filesystems..

* Remove unmountPending variable; Unmount now removes all known inodes
  directly, removing reliance on eventual consistency.

* Document Mount(), Unmount()

* Forced forgets (eg. when fusermount -u is called) will also cause
  FileSystem.Unmount() to be called.

* Changes to accomodate new conventions in ZipFs/UnionFs

* Configure MultiZipFs using symlinks rather than catching file
  writes.

* Remove return from Mount() api call.

openedFile.

To make this work, add OpenCountMutex to inode.  This fixes a race
condition in considerDropInode().

conversion.

FileSystemConnector.

Test for this in ZipFs.

This fixes a potential race conditions in
FileSystemConnector.internalLookupWithNode, which did not take the
treeLock.

store.

Add test.

Finding the workspace for a given backing store is a common operation
that requires readdir, and should not be expensive.

Further simplifies code.

Deal with nonexistent mountpoints both in mount and umount