As reported by @jjourdan: framing now no longer back tracks on whether
to strip laters or not. When framing below a later, we now only make it
strip laters of the head of the frame.

We now use the `Maybe` prefix as also used for `Frame`: it indicates
whether progress has been made by stripping of a later or not.

We already supported framing under wands.

For example, framing `P` in `(P ∨ Q) ∗ R` now succeeds and turns the goal into `R`.

This is to make sure that e.g. `//` in `iEval (rewrite .. // ..)` does
not immediately close the goal by reflexivity.

This should fix iris-examples.

This used to be done by using `ElimModal` in backwards direction. Having
a separate type class for this gets rid of some hacks:

- Both `Hint Mode`s in forward and backwards direction for `ElimModal`.
- Weird type class precedence hacks to make sure the right instance is picked.
  These were needed because using `ElimModal` in backwards direction caused
  ambiguity.

This was an oversight in !63.

It used to be an inline pattern match.

This also restores compatibility with Coq 8.6.1.

The proof mode now explicitly keeps track of anonymous hypotheses (i.e.
hypotheses that are introduced by the introduction pattern `?`). Consider:

  Lemma foo {M} (P Q R : uPred M) : P -∗ (Q ∗ R) -∗ Q ∗ P.
  Proof. iIntros "? [H ?]". iFrame "H". iFrame. Qed.

After the `iIntros`, the goal will be:

  _ : P
  "H" : Q
  _ : R
  --------------------------------------∗
  Q ∗ P

Anonymous hypotheses are displayed in a special way (`_ : P`). An important
property of the new anonymous hypotheses is that it is no longer possible to
refer to them by name, whereas before, anonymous hypotheses were given some
arbitrary fresh name (typically prefixed by `~`).

Note tactics can still operate on these anonymous hypotheses. For example, both
`iFrame` and `iAssumption`, as well as the symbolic execution tactics, will
use them. The only thing that is not possible is to refer to them yourself,
for example, in an introduction, specialization or selection pattern.

Advantages of the new approach:

- Proofs become more robust as one cannot accidentally refer to anonymous
  hypotheses by their fresh name.
- Fresh name generation becomes considerably easier. Since anonymous hypotheses
  are internally represented by natural numbers (of type `N`), we can just fold
  over the hypotheses and take the max plus one. This thus solve issue #101.