perf(save_tt): vectorize style resolution and typst CSS builder

[EinMaulwurf]

NOTE: This PR and the code changes are entirely "vibe coded" using GLM-5.2. So if that is a dealbreaker, please reject, I do understand. But this is a real-world issue I'm facing when generating lots of large typst-tables with lots of formatting. So I hope I can help. If wanted, I can provide further explanation and benchmark/test scripts.

---

Two phases of build_tt() dominate save_tt() time on tables with many style_tt() calls (heat-map backgrounds, per-column alignment, etc.):

1. Resolving @Style into the rectangular @style_other/@style_lines grids. apply_style_to_rect() re-scanned the full cell grid for every style entry (O(N * cells * props)) and append_lines_to_rect() used rbind() in a loop (O(N^2)).
2. The typst style_eval() backend, where process_typst_other_styles() built CSS one cell at a time via regex typst_insert_field() calls, and typst_apply_styles() grew style_dict_entries with c().

Replace with batched/vectorized equivalents:

• expand_style.R: new resolve_styles_batch() builds a single (i,j)->linear-index map and walks each property once. Old per-row functions kept for backwards compatibility.
• build_tt.R: calls resolve_styles_batch() instead of the per-row loop.
• typst_style.R: new .typst_build_css_vectorized() emits all key:value fragments in a single pass with matching output format; typst_apply_styles() uses a pre-allocated list + unlist().

Output is byte-identical across 6 fixtures x 4 formats (typst, html, latex, markdown) and the full tinytest suite passes.

Benchmarks (median of 5 reps, 60x12 table with 170 style_tt() calls):
typst: 1.091s -> 0.703s (1.55x)
html: 1.442s -> 1.151s (1.25x)
Style resolution phase: 0.451s -> 0.137s (3.3x)

[vincentarelbundock]

Thanks a lot for taking the time to put this together! This seems like a clear improvement. I've added a few other things and a thank you in the NEWS.

Merged.