rustc_parse: Better spans for synthesized token streams I think using the nonterminal span for synthesizing its tokens is a better approximation than using `DUMMY_SP` or the attribute span like #79472 did in `expand.rs`. r? `@Aaron1011`