Eglot: partial fix for middle-of-symbol completions

Eglot: partial fix for middle-of-symbol completions

* lisp/progmodes/eglot.el (eglot-completion-at-point): Fix
completion reversion in :exit-function.

In a rust-ts-mode buffer such as this main.rs file

  fn main() {
    let v: usize = 1;
    v.c<cursor-here>1234.1234567890
  }

the server wants to edit the line to read, after C-M-i and selecting
"count_ones"

    v.count_ones<cursor-here>.1234567890

But it couldn't apply the edit to the correct initial state because
that state wasn't correctly restored.  This commit fixes that.

However, if the initial state is

    v.count_on1234.1234567890

then completion still fails, because the 'try-completion' call in
eglot-completion-at-point will just return complete to "count_ones"
and Emacs doesn't consider this a completion "exit", so it'll
completely ignore the exit function.

I think 'try-completion' (and 'test-completion') simply can't be used
here (for one, they obey styles, and styles are off-limits in LSP),
but I'll leave that for another commit.

Github-reference: https://github.com/joaotavora/eglot/issues/1339