This is inspired by ssr, and makes unification faster if it goes
right-to-left.

See https://sympa.inria.fr/sympa/arc/ssreflect/2013-11/msg00010.html

This way, we can use eapply instead of class_apply, which is used
when the instances are defined using the Instance command. It seems
that eapply is stronger as class_apply, and as such solves some issues
when canonical structures have type class parameters, for example:

  Goal Op (option (dec_agree nat)). apply _.

This failed, but is fixed by this commit.

Based on an idea and WIP commits of J-H. Jourdan: the core of a CMRA
A is now a partial function A → option A.

TODO: define sum CMRA
TODO: remove one shot CMRA and define it in terms of sum

- Make the carrier argument of the constructors for the canonical structures
  cofeT and cmraT explicit. This way we make sure the carrier is properly
  exposed, instead of some alias of the carrier.
- Make derived constructions (such as discreteC and discreteR) notations
  instead of definitions. This is yet again to make sure that the carrier is
  properly exposed.
- Turn DRA into a canonical structure (it used to be a type class).

This fixes some issues, notably it fixes some broken rewrites in algebra/sts
and it makes canonical structures work properly with dec_agree.

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

Thanks to Amin Timany for the suggestion.

So, only use the type class for contractive functors.

Add both non-expansive and contractive functors, and bundle them for the general Iris instance as well as the global functor construction

This allows us to move the \later in the user-defined functor to any place we want.
In particular, we can now have "\later (iProp -> iProp)" in the ghost CMRA.

Since functor instances are just used as combinators, there is really
no need for functors that are not contractive.

This cleans up some ad-hoc stuff and prepares for a generalization
of saved propositions.

This replaces f_equiv and solve_proper with our own, hopefully better, versions