rio-vt: intern extras by content, copy-on-write attach

[raphamorim]

Workstream 0b of the mode-2027 scoping — and, like 0c/0d, it turned out to fix live bugs on main, not just prepare for clustering.

Two bugs on main

1. Combining marks leak across hyperlink spans. Every cell written under one OSC 8 open references the same extras slot (the cursor template holds one id). The zero-width attach path mutated that slot in place — so typing a combining mark inside a hyperlink attached the mark to every cell of the span. With rioterm: shape attached combining marks with their base glyph #1850 (marks now shape), that would render visibly on every cell.
2. The extras id space is a silent cliff. Slots cap at u16::MAX and overflow logs + drops the extras (grid/mod.rs:179). Every occurrence of every combining sequence burned a fresh slot toward it.

The change

• Content-hash interning: ExtrasTable::alloc reuses the slot of identical content (Extras gains Hash; Hyperlink already had it). Repeated emoji/marks cost one slot total — the mode-2027 prerequisite, where extras become the common case. Sweep/free/clear purge the lookup so stale content can't resolve to a dead slot.
• Copy-on-write attach: shared slots make in-place mutation illegal; the combining-mark path now clones, extends, and re-interns, giving the marked cell its own (marks + hyperlink) slot while its neighbors keep theirs. Fixes bug 1.
• Span detection by hyperlink, not slot id (hints + hover walk): a marked cell inside a link legitimately holds a different slot while belonging to the same span. Hyperlink is Arc-backed; comparison is content equality, matching OSC 8 id= semantics. This was subtly wrong before interning too — bug 1's flip side, a span splitting at a marked cell.

One test contract updated: vs16_at_last_column_preserves_base_extras pinned slot-id identity across the VS16 wrap, which encoded the in-place-mutation behavior; it now asserts the content ([ZWJ, FE0F] following the base), which is what consumers read.

Cost

Interning hashes content on alloc only — allocs happen per OSC 8 open and per combining-mark attach, not per cell or per frame. The renderer's per-frame ExtrasMap snapshot is untouched (and gets smaller with dedup).

Verification

Three new tests: intern-by-content (same marks share a slot, different marks don't), the hyperlink-leak regression (mark stays on one cell, neighbor keeps the pristine link slot, both keep the hyperlink), and lookup purging across reclaim. Full suites: 432 rio-vt, 46 librio, 172 rioterm; clippy/fmt clean.

[raphamorim]

Folded into the consolidated unicode-foundation PR (all four workstream-0 pieces in one branch for testability, commits preserved).