Markd::Rule::Paragraph

Upstream re-sliced the rest of the paragraph after every reference definition (container.text = container.text.byte_slice(pos)), so a paragraph of N consecutive definitions cost O(N²) in copying — and the inline lexer's per-paragraph caches, keyed on the text object, were rebuilt for every slice. Walk one string with an offset instead and slice once at the end; each definition is read by Inline#hwaro_reference_at exactly as reference read the slice.