Rework and correct major part of xref glue code

Rework and correct major part of xref glue code

See comments of https://github.com/joaotavora/eglot/pull/314.  Up to
now, xref-backend-indentifier-completion-table was a gross hack that
only worked sometimes.  It relied on some fugly gymnastics to cache a
response from :textDocument/documentSymbol and somehow used that
information to build a completion table.  But it doesn't work well.

Summarily, LSP doesn't lend itself well to the xref interface of
prompting for an arbitrary identifier and then go look for whichever
type of references of that identifier.  All the LSP
:textDocument/{definition,references,implementation,...} methods
expect to know the exact context of the search the user is about to
perform, in the form of a document location.  That conflicts with the
xref "arbitrary string" requirement.

Therefore, the slightly limited, but much more correct way, for Eglot
to function is to override the user's preference of
xref-prompt-for-identifier, temporarily setting it to nil in
eglot--managed-mode (ideally, though, xref-prompt-for-identifier
should be a function of the backend.)

Later on, a possibly better behaved identifier completion table can be
built on top of the :workspace/symbol LSP method.

* eglot.el (xref-backend-identifier-at-point): Rewrite.
(eglot--lsp-xrefs-for-method): New helper.
(eglot--lsp-xref-helper): Use eglot--lsp-xrefs-for-method.
(eglot--xref-definitions-method): Delete.
(eglot--lsp-xref-refs): New variable.
(xref-backend-references, xref-backend-definitions): Use
eglot--lsp-xrefs-for-method.
(eglot--managed-mode): Set xref-prompt-for-identifier
to nil.
(eglot--xref-reset-known-symbols, eglot--xref-known-symbols): Delete
(xref-backend-identifier-completion-table): Nullify.
(eglot-find-declaration, eglot-find-implementation)
(eglot-find-typeDefinition): Use eglot--lsp-xref-helper.