fix bug with rootObj optimization in observed sets

[jmesserly]

while investigating https://github.com/Polymer/polymer-dev/issues/101, I found two bugs with the rootObj optimization in observed sets:

• Models can have cycles, therefore, we may encounter the "root object" further down the path but not remember this fact. This leads to incorrectly failing to fire a change notification.
• At first, my new test only failed when run by itself, but passed in the full suite. This is because of a second bug: lastObservedSet can sometimes be the set we just closed. This leads to confusion about whether the observed set was closed or not, ultimately leaving it pointing at the wrong rootObj in its closure. I tried to mitigate this in two ways: instead of this.object and rootObj both tracking the same value, this is now handled explicitly in a setter to ensure they stay in sync. Also, if we close an observer make sure it is not still listed as the (alive) lastObservedSet.

[jmesserly]

@rafaelw and/or @arv mind taking a look?

[jmesserly]

Oh, by the way ... thoughts on sending this as a pull request instead of rietveld? Github has side-by-side view now so I've been pretty happy with it as a review tool, but curious what y'all think.

[arv]

LGTM but let @rafaelw have a look to since I haven't fully grok the implication yet.

[jmesserly]

friendly ping for this one & https://github.com/Polymer/observe-js/pull/74

(btw, thanks for the test comments @arv, will fix those before landing.)

[jmesserly]

going to land this as my other CL sort of depends on it (if I want to make the fix more general as suggested in https://github.com/Polymer/observe-js/pull/74#issuecomment-56866498). I'm pretty confident in this fix, but the main thing that worried me was a style issue: using a getter/setter for rootObject. I wasn't sure if using getters/setters is typical or if I should try and change how the newObservedSet closure is working (e.g. perhaps convert ObservedSet to the style of the other types in this file, where it puts the methods on the prototype instead of the instance?). Very happy to iterate on this though, so feel free to add post-submit comments.