If the BI is not affine, this should not happen, as it may lead to
information loss.

This commit fixes issue #190.

move test suite out of theories/ so it does not get installed; also check output of test suite so that we can test printing

It now turns the goal into `P` and `<pers> Q`, which is dual to
`iDestruct`, which turns `P ∧ <pers> Q` into `P` and `□ Q`.

This follows the proof at https://en.wikipedia.org/wiki/L%C3%B6b's_theorem#Proof_of_L%C3%B6b's_theorem

This fixes issue #182.

Fixes #176

This reverts commit daaabf41.  We have tests for
testcase, we shouldn't depend on using arcane features in our definitions for
this purpose.  Also, this exposes that there still is a bug somewhere (see the
FIXME).

We do this in two ways:

- Use `notypeclasses refine` instead of `eapply`, to avoid type class
  search being called arbitrary.
- Use `typeclasses eauto` instead of `apply _`, to avoid type class
  search being called on unrelated evars.

I mainly tried this for `iSpecialize` and friends; this same remains to
be done for all other tactics.

This commit also makes partial progress w.r.t. issue #135.

This requires changing the Hint Mode of the [Frame] type class because it should not fail if its parameter is an evar, but instantiate it instead. In order to prevent all the other instances of [Frame] to intantiate this evar themselves, we create a new type class [KnwonFrame], which corresponds to the old behavior.

They are split into:
1- KnownMakeXXX , which only works if the parameter is not an evar. Hence, it will never force this evar to becomes e.g., emp or True.
2- MakeXXX, which works even if this is an evar, but it only has instances that will not instanciate arbitrarilly this evar.

As requested by @jtassaro.

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.