fix: keep grapheme clusters whole in block cursor and gradient writer

Apply review findings: the textarea block-cursor overlay dropped only the
final rune (splitting a flag or ZWJ family at a line end), and
Buffer.SetStringGradient was the last per-rune writer (over-counting clusters
and splitting glyphs). Both now operate on whole clusters. Also cross-document
the deliberate graphemeExtend vs isZeroWidthRune split.