Optimize resolve_xref to avoid O(n^2) iteration

Optimize resolve_xref to avoid O(n^2) iteration

Formerly resolve_xref() in Sphinx extension called get_objects() from
Sphinx API which subsequently iterated the whole list of objects,
causing single iteration per single reference, which is essentially.
O(n^2). Avoid using get_objects() and access internal dictionary
directly intead.

The docs build time was still dominated by other factors but speedup is
about 10 % on my machine.

(cherry picked from commit 8a5f20121ea1fc5375fab0ba777d36a5605e3763)