The case of lam and case expressions that before required the terms to be
well-typed now require terms to be closed.

Separated definition context and context refinement from soundness_binary file.

Squashed commit of the following:

commit a8d2dd620df2fe8531b590811b7f08d2bc1289b4
Author: Amin Timany <amintimany@gmail.com>
Date:   Sun May 29 13:54:07 2016 +0200

    Prove refinement of fine/coarse-grained stack

commit 6347ef920581b4f21b5dfa74d288afcf482c9b50
Author: Amin Timany <amintimany@gmail.com>
Date:   Sun May 29 01:37:23 2016 +0200

    Backup

commit 39552d8055f55458c9515e629707d496e26e92b7
Author: Amin Timany <amintimany@gmail.com>
Date:   Sat May 28 22:40:02 2016 +0200

    Backup

This is to avoid confusion when Coq loads the module.

It now produces a lambda that when evaluated with a value runs the underlying
lambda that value after acquiring the lock. As before, the lock is released
afterwards.

It now returns the value of the expression evaluated.

We used to have:

(λ x, b) a ->β b[a/x]

But now we have:

(λ f x, b) a ->β b[a/x, (λ f x, b)/f]

There is one admit that needs to be taken care of. The admitted case is validity
of a monoid element of iprod. At the moment iris doesn't seem to have necessary
lemmas to prove this easily.

With this change, recursive types don't use the context for type variables.
This allows us to assume that all types in typing context are universally
quantified, i.e., they are all polymorphic types.