feat(renderer): font ligature support via GSUB shaping + segmentation cache

[patrick-andrew-anchor]

Programming ligatures (=>, ->, !=, ===, <==>, ...) now render in the dynamic atlas for fonts that ship ligature tables (Fira Code, JetBrains Mono, Cascadia Code, Monaspace Neon).

The dynamic atlas rasterized one grapheme per cell, so the font shaper never saw adjacent codepoints and ligatures never formed. This adds a ligature shaper plus N-cell glyph support end to end:

• beamterm-core (new ligatures feature):
  • shaper.rs: rustybuzz-based detection. Compares each shaped glyph to the nominal cmap glyph, so it detects the calt "spacer" approach (Fira Code et al. keep glyph count == char count) as well as classic GSUB ligature merges.
  • GlyphSlot::Ligature(id, cells) + cell_span(); size-classed ligature pools (widths 3..=8) in the glyph cache with O(1) alloc + LRU eviction. Two-cell ligatures continue to use the existing wide path.
  • dynamic atlas: generic split_glyph_n + N consecutive slot uploads; texture layers derived from the region layout so they can't drift.
  • terminal grid: 2-cell placement generalized to N cells across all update paths; segment_run + place_ligature helpers.
• beamterm-renderer:
  • canvas rasterizer sizes each glyph to cell_w * unicode-width, so ligature substrings render at their full width.
  • BeamtermRenderer.setFontBytes(Uint8Array) builds the shaper from raw sfnt bytes; Batch.text segments runs into ligature glyphs. Ligatures activate automatically when the supplied font advertises them.

The shaper only detects/segments; the browser canvas still rasterizes, preserving color emoji and font fallback. WOFF/WOFF2 must be decompressed to sfnt before setFontBytes (documented in js/README.md).

Testing:
With this change and my client support, I can render ligatures:

user@pandrew-dev:~$ printf '%s\n' '-> => != == === !== <==> --> |> :: // /* */ >>= <=>'
  -> => != == === !== <==> --> |> :: // /* */ >>= <=>"

shows as

---

Follow-up: memoize ligature run segmentation

This PR also includes a perf commit on top of the feature.

Shaper::segment() built a rustybuzz Face from the raw font bytes and ran the shaper for every text run on every call. The renderer re-shapes the whole screen each frame, so a static screen was re-segmented ~60×/s — measured at ~38ms/frame of pure shaping on a full screen (render p50 48.8ms with ligatures on vs ~10.5ms off).

segment() is now memoized in an LRU keyed on the run text. Segmentation depends only on the characters and the font, and a font change constructs a fresh Shaper (hence a fresh cache), so no explicit invalidation is needed. A static screen pays the shaping cost once; repeated runs become an O(len) map lookup. lru was already a dependency.

Adds a cache-correctness test asserting the memoized path returns segments identical to the uncached path on both miss and hit (gated on BEAMTERM_LIGATURE_TEST_FONT). Confined to beamterm-core/src/gl/shaper.rs; no new dependencies.