Fix eglot-completion-at-point to work with bare completion-at-point

Fix eglot-completion-at-point to work with bare completion-at-point

Fixes https://github.com/joaotavora/eglot/issues/313, fixes https://github.com/joaotavora/eglot/issues/311, fixes https://github.com/joaotavora/eglot/issues/279

As is well known, LSP's and Emacs's completion mechanics don't fit
very well together, mostly because Emacs expects completion to be a
something of a pure function of a string argument and LSP treats as a
function of a concrete buffer position.

A further complication arises because some completion frontends like
"bare" completion-at-point make Emacs modify the buffer's contents
during the completion process, while other (notably company-mode)
don't do that.  Thus, 'eglot-completion-at-point' must take extra care
to answer to the questions listed in the "(elisp)Programmed
Completion" info node based on its (quite hacky) "completions" local
var and _not_ based on the intermediate buffer contents.  That var is
also used to cache the last LSP response and allow the :exit-function
callback to retrieve much more than just the completion text in

In yet another related problem, :exit-function won't be called at all
with completion-at-point if the completion table doesn't answer
properly to test-completion.  A previous use of
completion-table-dynamic was found to be unsuitable here: we must
answer all the requests separately.

* eglot.el (eglot-completion-at-point): Rework.