iSpecialize and iDestruct.

These tactics now all take an iTrm, which is a tuple consisting of
a.) a lemma or name of a hypotheses b.) arguments to instantiate
c.) a specialization pattern.

These tactics are superfluous:
- iPure H as pat => iDestruct H as pat
- iPersistent H => iSpecialize H "!"

It is no longer triggered when posing [P ⊢ Q] with [P] an evar. This, for
example, makes sure that iApply pvs_intro works, which failed before.

This reverts commit 327acdbd. I thought we could
get something with a contractive "un-injection" funtion, but that turns out to
be impossible.

Now, it bases the type the quantifier ranges over on the goal, instead
of the witness. This works better when dealing with witnesses involving
type class constraints.

- It can now also frame under later.
- Better treatment of evars, it now won't end up in loops whenever the goal
  involves sub-formulas ?P and it trying to apply all framing rules eagerly.
- It no longer delta expands while framing.
- Better clean up of True sub-formulas after a successful frame. For example,
  framing "P" in "▷ ▷ P ★ Q" yields just "Q" instead of "▷ True ★ Q" or so.

Par is now defined as an expression of type [∀ X, expr X] (instead of a
value) and we prove that it is stable under weakening and substitution.

It is just a test case and not really part of the barrier library.

This gets rid of the (ambiguous) notation %l, because we can declare
LitLoc as a coercion. It also shortens the code.