proof-state-change-pre-hook happens earlier than
proof-state-change-hook, i.e. before proof-done-advancing. This should
be used to register information in the currently processed span before
proof-done-advancing classifies it.

Historically PG design was to gather these information during
proof-done-advancing (or in its hook called at the end) by just
looking at the command statement. But it is often useful to look at
the output (messages and/or prompt) to gather more accurate
information. Some of this information may be needed DURING
proof-done-advancing. Hence this early hook.

Use lexical-binding in a few files where it was easy.
Don't require `proof-compat` when it's not used.

* coq/coq-db.el: Use lexical-binding.

* coq/coq-system.el: Use lexical-binding.
(coq--extract-prog-args): Use concatenated-args rather than recomputing it.

* coq/coq.el: Require `span` to silence some warnings.

* generic/pg-user.el: Use lexical-binding.
(complete, add-completion, completion-min-length): Silence warnings.

* generic/pg-xml.el: Use lexical-binding.
(pg-xml-string-of): Prefer mapconcat to reduce+concat.

* generic/proof-depends.el: Use lexical-binding.
(proof-dep-split-deps): Use `push`.

* generic/proof-shell.el: Require `span` to silence some warnings.
(proof-shell-invisible-command): Don't use lexical-let just to build
a wasteful η-redex!

* lib/holes.el: Use lexical-binding.
Remove redundant :group args.

* lib/span.el: Use lexical-binding.
(span-read-only-hook): Use user-error.
(span-raise): Remove, unused.

The problem is that loading pg-custom runs a bunch of defpgcustom, with no
current proof assistant.  Then when coq or easycrypt calls
proof-ready-for-assistant, pg-custom isn't loaded again.