Alternative implementation of cursors

[ept]

This PR provides an implementation of cursors, which allow a particular element of a list, or a particular character in an Automerge.Text object, to be referenced from elsewhere in the same document. The API involves a new Automerge.Cursor class, and it can be used like this for lists:

let doc = Automerge.change(Automerge.init(), doc => {
  doc.list = [1, 2, 3]
  doc.cursor = new Automerge.Cursor(doc.list, 2)
})
doc.cursor.index // returns 2

and like this for Text:

let doc = Automerge.change(Automerge.init(), doc => {
  doc.text = new Automerge.Text(['a', 'b', 'c'])
  doc.cursor = doc.text.getCursorAt(2)
})
doc.cursor.index // returns 2

The key thing about a cursor is that its index automatically updates when the list/text it references is changed. In the examples above, the index updates to 3 if you insert a new element before the cursor position, and it updates to 1 if you delete an element before the cursor position. If you delete the element that the cursor points to, the index is set to the index of the closest preceding element that has not been deleted (or -1 if there is no such element that has not been deleted).

The implementation of cursors on #307 requires a direct call from the frontend into the backend, which is inconvenient when frontend and backend are on different threads. This PR is an alternative implementation that is integrated with the existing frontend-backend protocol, making it compatible with multiple threads. This PR also includes a way for cursors to be efficiently represented in the binary storage format for changes and documents. A downside of this approach compared to #307 is that this PR requires cursors to be stored within the same Automerge document; it is not possible to store them somewhere out-of-band, or in a separate document.

[ept]

This PR is now mostly complete and working: cursors can be created and stored, and indexes are automatically recomputed as needed. There is one known issue: Automerge.load(Automerge.store(doc)) currently throws an exception when loading a document that contains cursors (there's a skipped test for this). Storing and loading a document should work if it is done as Automerge.applyChanges(Automerge.init(), Automerge.getAllChanges(doc)). I will try to fix this soonish, but otherwise this feature is ready for people to start playing with.