storage has a failing abortVersion(), so it's mostly a checkpoint.

Not yet tested: commitVersion(), transactionalUndo(), undoLog().

Not yet implemented: history(), pack()

Specific changes:

    - Table txnOids renamed to txnoids everywhere.

    - Define a module global `zero' constant and use (via global)
      where ever appropriate (should be named ZERO but that's ugly).

    - _finish(): Only boost the refcount of objects referenced by the
      pickle when we've actually got a new pickle.  Also, add a note
      about the currentVersions table, which will contain multiple
      copies of vid/oid records.  It's easier to weed them out later
      than to weed them out here.  One of the places you need to weed
      them all out is at the end of this method, in the
      `delete-a-version' branch.

    - abortVersion(): Record oids found by scanning the
      currentVersions table in a set, to account for the possible
      multiple vid/oid records.  Fix some typos.  When calling
      write_nonversion_object() on the CommitLog, the prevrevid ought
      to be the revid of the current object revision, for undo
      purposes.

    - load(): Where we previously had a comment that asked whether we
      should assert that the version is empty (after we've checked
      everything else), now we really check that and raise a
      VersionError if not.

    - loadSerial(): Fix a typo.

    - store(): Inside the old-revision-id-is-true test, if the old vid
      is not zero and it is equal to the new vid, then the non-version
      revision of the object is the old non-version revision (added
      the else: clause).

    - transactionalUndo(): Implemented but not yet tested.

    - undo(): Removed.

    - versionEmpty(): Semantics are to return true if an unknown
      version is given.  As an extra bonus, if version is empty
      (i.e. the non-version) check the serials table.