Parser: extend spans during Pratt parsing for better error reports.

[Alonely0]

This is currently done only in some places; the idea is to keep an "anchor" of where the subexpression's span starts, and when reporting an error, to replace the problematic token's span start with it.