Variable class (like tf.Variable) that supports eager mode

I've been looking at adding a tf.Variable like class, that would work in eager and graph mode (the variable op doesn't work in eager), and register itself in the execution environment. It's not terribly useful until we get eager gradient support, but it's not dependent on it either so there's no reason to wait.

Eager mode is easy, just have a field. For graph, we can store a Variable op, assign to it, but return/expose the return of the assign (it's the new value). The assigns can take the last assign as a control dep.

I've got a draft here I can PR, but I'm not sure how much the type system refactor would affect it. At a minimum I don't see a reason to merge before #160.

Start by reviewing the linked rn_variable draft and the type system refactor discussed in issue #160. Check how the proposed variable would represent eager and graph-mode state and register itself in the execution environment. Done means a tf.Variable-like class supports both modes without depending on eager gradient support.