Prevent stale servers when using eglot-extend-to-xref

Prevent stale servers when using eglot-extend-to-xref

A weak-valued hash-table is not enough to guarantee that a reference
to a zombie server in eglot--servers-by-xrefed-file variable won't
survive long enough to confuse the next call to eglot--current-server
in some buffers.

So, before this fix it was common to get "Process EGLOT ... not
running" errors if some xref-extended buffers (like system libraries)
were open and M-x eglot-reconnect was issued.  This should be
prevented now.

Note however, that even after this the eglot-extend-to-xref logic is
still flawed.  For example, if a buffer for the xref-extended buffer
happens to be already visited by the time M-. is issued to navigate to
it, Eglot won't be activated.  A half-decent workaround is to kill the
buffer and re-visit it.

* lisp/progmodes/eglot.el (eglot--servers-by-xrefed-file): Move
up.
(eglot--on-shutdown): Make sure to cleanup
eglot--servers-by-xrefed-file.