[clangd] Don't trim xrefs references if we overran the limit

This preserves all the results we've processed already rather than
throwing them away in the end.
It has some performance implications on the edge cases, in the worst case we
might issue 1 relations and 2 xrefs requests in extra to deduce `HasMore`
correctly.

Fixes https://github.com/clangd/clangd/issues/204.

Differential Revision: https://reviews.llvm.org/D116043