fix(rio-grid): prevent RTL glyphs from collapsing into one cell

[SagiMedina]

RTL glyphs can collapse onto one terminal cell because the grid emitter only advances its cluster-to-cell cursor. For example, source clusters [3, 2, 1, 0] currently produce cell indices [3, 3, 3, 3]. This change maps them to [3, 2, 1, 0], retaining each glyph's logical terminal cell and its foreground color.

The mapping keeps the existing forward walk for ascending clusters and uses a binary search when a glyph belongs to an earlier cell. It supports direction changes within a run without classifying the whole run by its first and last clusters. Both platform offset conventions (UTF-16 on macOS, UTF-8 elsewhere) and the existing wide-cell column mapping are preserved.

This fixes cell attribution for RTL scripts, including Hebrew, Arabic, and Persian. It preserves Rio's logical cell order; it does not implement Unicode BiDi visual reordering, proportional RTL layout, or new combining-mark positioning. Those need coordinated cursor, selection, and mouse mappings.

Related: #1877, which identified the same forward-only mapping bug. This is a narrower alternative that retains terminal cell positions. The broader BiDi feature request #1257 remains open.

Validation on macOS arm64 with Rust 1.96.1:

• Five regression tests failed against the original mapping, including real CoreText shaping and the production row emitter, and passed after the fix.
• All 24 rio-grid tests pass. Coverage includes RTL and mixed-direction clusters, UTF-8/UTF-16 offsets, marks and ligatures, actual Hebrew/Arabic/Persian shaping, and row emission into a CPU atlas with per-cell colors and cursor/selection run breaks.
• cargo test --features wgpu --locked passes across the full workspace (1374 tests passed, 11 ignored, including doc tests).
• cargo fmt --all -- --check and cargo clippy --all-targets --all-features --locked pass.

Interactive GPU rendering and Android/tmux handoff have not been manually tested with this patch.