This fixes the problem in iris/iris!275 (comment 42062)

This fixes some issues in GPS and RustBelt-relaxed, where the old
unification algorithm is now too weak.

This is needed to fix RustBelt-related. Since type classes are no longer
over-eagerly resolved, the `biIndex` now sometimes contains an evar.

Also, rewrite `iIntoEmpValid`. Now, instead of using Ltac to traverse
the type of the term and generate goals for the premises, we repeatedly
apply a series of lemmas. This has the advantage that it works up to
convertability, and we no longer need the `eval ...` hacks.

To prepare for https://github.com/coq/coq/pull/10762

The space should not be there and was added in oversight. This also provides
forwards compatibility with https://github.com/coq/coq/pull/10832.

Fix issue pointed out by @tchajed in iris/iris!330 (comment 41322)

Since Coq 8.10, `move => {}e` means `move => {e}e`.
For the backward-compatible syntax and discussion, see
https://github.com/coq/coq/issues/10550#issuecomment-542397265.

There is no need to include the `(∃ P', □ ▷ (P ↔️ P')  ...` since we
get closure under `▷ □ ↔️` from regular invariants.

Due to the new semantic invariants (!319) we no longer need to
close the model (i.e. `inv_def`) to be contractive, the semantic
invariant definition (i.e. `inv`) is already contractive.