fix(render): evaluate multi-input Type 0 tint transforms for DeviceN

[yfedoseev]

Summary

Closes #859.

Corrected root cause. The issue title (and its own follow-up triage comment) suspected Type 4 (PostScript calculator) tint transforms. That's stale — the actual repro file (devicen.pdf, a pdf.js corpus fixture) has an incremental update, so the live content stream (verified with pikepdf, which resolves the xref/generation chain correctly — naive byte-search over the raw file finds a stale first revision) uses a Type 0 (sampled) tint transform on a 3-channel /DeviceN [/Red /Green /Blue] color space, /Domain [0 1 0 1 0 1], /Size [13 13 13] — genuinely 3-dimensional.

resolve_separation_or_devicen in src/rendering/resolution/color.rs only ever forwarded components[0] (the first scn operand) to the Type 0/3 evaluator, discarding every other input channel. That's invisible for the common single-channel Separation case, but this DeviceN space always supplies 3 components. The 1-D-only sampled-function evaluator (evaluate_type0_sampled) then rejected the request (Size.len() == 3 != 1) and fell back to the long-standing gray = 1 - components[0] heuristic. Shape 1's first component was 1.0 → black. Shapes 2 and 3's first component was 0.0 → pure white, invisible against the white page — the "missing shapes" symptom. Type 4 dispatch (already correctly threading the full component vector, from #849) was never the problem.

Fix. Generalized evaluate_type0_sampled to N-dimensional multilinear interpolation across the 2^N sample-grid corners (ISO 32000-1 §7.10.2's general algorithm, of which the prior 1-D linear interpolation is the N=1 special case), with a MAX_SAMPLED_FUNCTION_DIMS = 8 guard against a pathological /Size array. Threaded the full components slice through for Type 0 (Type 2/3 stay single-input, matching their spec definition).

Test plan

• cargo build, cargo fmt --check, cargo clippy --features rendering -- -D warnings — clean.
• cargo test --lib --features rendering rendering::resolution::color — 16/16 pass, including two new tests: type0_sampled_function_single_dimension_matches_prior_linear_interpolation (pins the N=1 case is byte-for-byte unchanged) and type0_sampled_function_two_dimensional_input_uses_all_components (hand-computed 2-D grid, asserts a non-first-channel-dominant input produces the correct interpolated result).
• cargo test --features rendering — full suite green.
• Corpus regression: corpus_sig over the full local corpus (3530 PDFs) — zero-diff, identical output between main and this branch. (Expected: corpus_sig measures text-extraction signatures; this change only affects rendering-time color resolution, so it's a true negative-control confirmation that nothing else moved, not a direct exercise of the fix.)
• Visual verification against the real repro file (devicen.pdf, not committed — CONTRIBUTING.md fixture policy): before, one black triangle + two invisible-white ones; after, three distinct green/blue/red triangles, matching poppler's reference rendering described in the issue.