DESTROY is only sent for fuseblk devices which we (probably) don't
support anyway.

Previously, a FORGET opcode could force a mount that was unused to be
forgotten.  This would cause the mounting filesystem to get into
inconsistent states: MultiZipFs and AutoUnionFs would believe the
mount is still live, but FilesystemConnector had already forgotten it.

The downside to this solution is that submount FileSystems will no
longer be unmounted if the root FileSystem is unmounted with
"fusermount -u".

Add a gotest benchmark running off a synthetic stat-only filesystem.

mount point.

Test for behavior of deletion cache drop too.

in NewUnionFsFromRoots().

This allows Termite to use other forms of caching.

* 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.