Instead, I have introduced a type class `Monoid` that is used by the big operators:

    Class Monoid {M : ofeT} (o : M → M → M) := {
      monoid_unit : M;
      monoid_ne : NonExpansive2 o;
      monoid_assoc : Assoc (≡) o;
      monoid_comm : Comm (≡) o;
      monoid_left_id : LeftId (≡) monoid_unit o;
      monoid_right_id : RightId (≡) monoid_unit o;
    }.

Note that the operation is an argument because we want to have multiple monoids over
the same type (for example, on `uPred`s we have monoids for `∗`, `∧`, and `∨`). However,
we do bundle the unit because:

- If we would not, the unit would appear explicitly in an implicit argument of the
  big operators, which confuses rewrite. By bundling the unit in the `Monoid` class
  it is hidden, and hence rewrite won't even see it.
- The unit is unique.

We could in principle have big ops over setoids instead of OFEs. However, since we do
not have a canonical structure for bundled setoids, I did not go that way.

Big ops over list with a cons reduce, hence these just follow
immediately from conversion.

This fixes issue #84.

This way, iSplit will work when one side is persistent.

This could lead to awkward loops, for example, when having:

- As goal `own γ c` with `c` persistent, one could keep on
  `iSplit`ting the goal. Especially in (semi-)automated proof
  scripts this is annoying as it easily leads to loops.
- When having a hypothesis `own γ c` with `c` persistent, one
  could keep on `iDestruct`ing it.

To that end, this commit removes the `IntoOp` and `FromOp` instances
for persistent CMRA elements. Instead, we changed the instances for
pairs, so that one, for example, can still split `(a ⋅ b, c)` with
`c` persistent.

This are useful as proofmode cannot always guess in which direction
it should use ⊣⊢.

This fixes issue #81.

The instances frame_big_sepL_cons and frame_big_sepL_app could be
applied repeatedly often when framing in [∗ list] k ↦ x ∈ ?e, Φ k x
when ?e an evar. This commit fixes this bug.

- Allow framing of persistent hypotheses below the always modality.
- Allow framing of persistent hypotheses in just one branch of a
  disjunction.

This has some advantages:

- Evaluation contexts behave like a proper "Huet's zipper", and thus:
  + We no longer need to reverse the list of evaluation context items in the
    `reshape_expr` tactic.
  + The `fill` function becomes tail-recursive.
- It gives rise to more definitional equalities in simulation proofs using
  binary logical relations proofs.

  In the case of binary logical relations, we simulate an expressions in some
  ambient context, i.e. `fill K e`. Now, whenever we reshape `e` by turning it
  into `fill K' e'`, we end up with `fill K (fill K' e')`. In order to use the
  rules for the expression that is being simulated, we need to turn
  `fill K (fill K' e')` into `fill K'' e'` for some `K'`. In case of the old
  `foldr`-based approach, we had to rewrite using the lemma `fill_app` to
  achieve that. However, in case of the old `foldl`-based `fill`, we have that
  `fill K (fill K' e')` is definitionally equal to `fill (K' ++ K) e'` provided
  that `K'` consists of a bunch of `cons`es (which is always the case, since we
  obtained `K'` by reshaping `e`).

Note that this change hardly affected `heap_lang`. Only the proof of
`atomic_correct` broke. I fixed this by proving a more general lemma
`ectxi_language_atomic` about `ectxi`-languages, which should have been there
in the first place.

- Support for a `//` modifier to close the goal using `done`.
- Support for framing in the `[#]` specialization pattern for
  persistent premises, i.e. `[# $H1 $H2]`
- Add new "auto framing patterns" `[$]`, `[# $]` and `>[$]` that
  will try to solve the premise by framing. Hypothesis that are
  not framed are carried over to the next goal.