Since [inG] ranges over [cmraT]s, using an [ucmraT]s results in
[ucmra_cmraR] coercions that slow down type checking.

This commit improves the compilation time of thread_local.v by 40%.

By using the global ghost maps instead of our own ones.

It is not non-expansive, so not a function we should use.

This generalization is surprisingly easy in Iris 3.0, so I could not
resist not doing it :).

This makes stuff more uniform and also removes the need for the [inGFs]
type class. Instead, there is now a type class [subG Σ1 Σ2] which expresses
that a list of functors [Σ1] is contained in [Σ2].

I do not know why we have to split the rewrite here, but it seems we do.

This is probably due to a bug in the rewrite of ssreflect 1.6 which has
been fixed in ssreflect master.

I cannot reproduce the error of the CI builder on my machine with the
same version of Coq (8.5pl2).

And make it Typeclasses Opaque to ensure that we indeed do not do
so using the proof mode.

Instead of having connectives pvs0 and pvs1 we now have one connective
pvs that is indexed by a Boolean.

Use it to prove that tests/barrier_client and tests/heap_lang are adequate.

Also make those for introduction and elimination more symmetric:

  !%   pure introduction         %        pure elimination
  !#   always introduction       #        always elimination
  !>   later introduction        > pat    timeless later elimination
  !==> view shift introduction   ==> pat  view shift elimination

This commit features:

- A simpler model. The recursive domain equation no longer involves a triple
  containing invariants, physical state and ghost state, but just ghost state.
  Invariants and physical state are encoded using (higher-order) ghost state.

- (Primitive) view shifts are formalized in the logic and all properties about
  it are proven in the logic instead of the model. Instead, the core logic
  features only a notion of raw view shifts which internalizing performing frame
  preserving updates.

- A better behaved notion of mask changing view shifts. In particular, we no
  longer have side-conditions on transitivity of view shifts, and we have a
  rule for introduction of mask changing view shifts |={E1,E2}=> P with
  E2 ⊆ E1 which allows to postpone performing a view shift.

- The weakest precondition connective is formalized in the logic using Banach's
  fixpoint. All properties about the connective are proven in the logic instead
  of directly in the model.

- Adequacy is proven in the logic and uses a primitive form of adequacy for
  uPred that only involves raw views shifts and laters.

Some remarks:

- I have removed binary view shifts. I did not see a way to describe all rules
  of the new mask changing view shifts using those.
- There is no longer the need for the notion of "frame shifting assertions" and
  these are thus removed. The rules for Hoare triples are thus also stated in
  terms of primitive view shifts.

TODO:

- Maybe rename primitive view shift into something more sensible
- Figure out a way to deal with closed proofs (see the commented out stuff in
  tests/heap_lang and tests/barrier_client).

1.) iDestruct is able turns  into two implications (because uPred_iff
    is (type classes) transparent).
2.) iApply only backtracks on turning P  Q into P → Q or Q → P when
    there are no future premises. This is not the case for
    'P  □ (P → False)'.