Add a transactional write operation that will automatically retry

[deontologician]

Leaving this to @mlucy to fill out details.

[mlucy]

The proposal is basically that you can write something like this:

collection.find(id).atomicReplace(function(obj) {
  obj.a = obj.a + 1
  return obj;
})

(Where the name could obviously be better.)

Every object that Horizon stores in RethinkDB will have a $reql_version$ field (we've already reserved the $reql_*$ namespace for internal use) that is set to a UUID, and every write operation issued by Horizon will also update that version to a new UUID. This version will never be passed to the user of the Horizon client library, we'll strip it out before objects reach them.

atomicReplace will be implemented like so:

1. Fetch the document, record its version V.
2. Call the function passed to atomicReplace, get a new document D.
3. Issue this query:

collection.get(id).replace {|row| 
  r.branch(row['$reql_version$'].eq(V), D.merge({'$reql_version$' => r.uuid()}), r.error("retry"))
}

1. If we get the "retry" error, go back to step 1. Otherwise we're done.

---

The advantage of this interface is that it lets you issue a write that will never be clobbered by another concurrent write. So in the example above, if you issue 5 of those increment operations, you know that a will be larger by exactly 5 when they all complete.