This reverts commit 78ba9509.

This commit implements a generic `iAlways` tactic that is not tied to
`persistently`, `affinely` and `plainly` but can be instantiated with a
variety of always-style modalities.

In order to plug in an always-style modality, one has to decide for both
the persistent and spatial what action should be performed upon introducing
the modality:

- Introduction is only allowed when the context is empty.
- Introduction is only allowed when all hypotheses satisfy some predicate
  `C : PROP → Prop` (where `C` should be a type class).
- Introduction will only keep the hypotheses that satisfy some predicate
  `C : PROP → Prop` (where `C` should be a type class).
- Introduction will clear the context.
- Introduction will keep the context as-if.

Formally, these actions correspond to the following inductive type:

```coq
Inductive always_intro_spec (PROP : bi) :=
  | AIEnvIsEmpty
  | AIEnvForall (C : PROP → Prop)
  | AIEnvFilter (C : PROP → Prop)
  | AIEnvClear
  | AIEnvId.
```

An always-style modality is then a record `always_modality` packing together the
modality with the laws it should satisfy to justify the given actions for both
contexts.

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.

We do not have a notation for `bi_affinely` either, so this is at
least consistent.

(□ P) now means (bi_bare (bi_persistently P)).

This is motivated by the fact that these two modalities are rarely
used separately.

In the case of an affine BI, we keep the □ notation. This means that a
bi_bare is inserted each time we use □. Hence, a few adaptations need
to be done in the proof mode class instances.

Whenever we iSpecialize something whose conclusion is persistent, we now have
to prove all the premises under the sink modality. This is strictly more powerful,
as we now have to use just some of the hypotheses to prove the premises, instead of
all.

This also applies to the introduction pattern `!#`. Both will now
introduce as many ■ or □ as possible. This behavior is consistent
with the dual, `#`, which also gets rid of as many ■ and □ modalities
as possible.

(All the later lemmas are now prefixed by later_, and dito for
laterN, and except_0).

The absence of this axiom has two consequences:

- We no longer have `■ (P ∗ Q) ⊢ ■ P ∗ ■ Q` and `□ (P ∗ Q) ⊢ □ P ∗ □ Q`,
  and as a result, separating conjunctions in the unrestricted/persistent
  context cannot be eliminated.
- When having `(P -∗ ⬕ Q) ∗ P`, we do not get `⬕ Q ∗ P`. In the proof
  mode this means when having:

    H1 : P -∗ ⬕ Q
    H2 : P

  We cannot say `iDestruct ("H1" with "H2") as "#H1"` and keep `H2`.

However, there is now a type class `PositiveBI PROP`, and when there is an
instance of this type class, one gets the above reasoning principle back.

TODO: Can we describe positivity of individual propositions instead of the
whole BI? That way, we would get the above reasoning principles even when
the BI is not positive, but the propositions involved are.