feat(document): a text edit that spans several runs

[andiwand]

🤖 Generated with Claude Code

Second of five; #872 is merged, so this now sits on main. See
docs/design/document-editing.md for the decisions.

What this one does

insertText puts a run beside one that is there, in the same parent so it
carries the same style, and removeElement takes an element and its subtree
away. With setText on each end, a selection that starts in one run and ends
in another replays as the two ends rewritten and what lay between removed.

Document::remove, Document::insert_text_before and insert_text_after are
the same three in the C++ API. They sit on the document rather than on a
handle: a handle says what an element holds, the document says what the tree
holds. An Element is an immutable handle, so restructuring the tree through
one would leave a handle naming something unreachable. Each refuses an element
of another document.

Created elements

An operation that creates one states a negative id for it, and a later
operation names it by the same number. Replay keeps a per-log map. That keeps
replay a pure function of the log — returning minted ids to the browser is the
round trip architecture A exists to avoid. A sign rather than a reserved range,
because odf already spends the top bit of ElementIdentifier on a positional
cell id.

The registry grows the links a structural edit needs

insert_sibling_after, insert_sibling_before and unlink_child. Until now
nothing built a tree except a parser reading forward, so append_child was the
whole surface. An unlinked element keeps its id and stops being reachable, so
an id an edit already handed out never names something else.

xml::TreeEditor is where the dom half lives, over any registry whose elements
carry a pugi::xml_node. odf, ooxml text and ooxml presentation all resolve an
id to a node, splice the subtree and fix up the links; only the tag names
differ, and those come from the nodes.

Read-only engines take a default that throws, rather than eight identical
bodies.

Fix

A .docx edit dropped the space at either end of a run: the w:t now states
xml:space="preserve" where the text needs it ([ECMA-376] Part 1 17.3.3.31). A
lone space is part of a string token, so the token type never said one was
there. Verified with headless LibreOffice on the saved package.