docs: ndarray/nalgebra onboarding, llms.txt conventions, and zero-copy interop

[shinaoka]

Fixes #1641

Rust-native onboarding surface: ndarray/nalgebra priors no longer fall into the silent-transpose trap, llms.txt is actionable at one fetch, and external faer/ndarray memory is wrapped zero-copy.

What changed

• New mapping page docs/getting-started/ndarray-nalgebra-mapping.md (peer of pytorch-jax-mapping.md): the column-major asymmetry (ndarray row-major dangerous, nalgebra column-major fine), crate-selection entry point, why the backend is an explicit constructed value + the construct-once-reuse idiom, an faer-vs-BLAS decision guide, .dot()→matmul arity across the three API tiers, Tensor vs TypedTensor, and a zero-copy-interop section stating that adopting tenferro kernels does not require migrating tensor ownership (with the measured +10 crates / +28 s cold build from [Feature]: Guarantee downstream faer and BLAS interop with runnable tutorials #1602).
• llms.txt: inline "Conventions you must know before writing code" block (the five conventions), republished skill-reference recipes near the top, warning-shaped descriptions for the memory-order / einsum / parallelism-and-caching trap pages, and a link back to the README as the canonical router.
• README: the same five-line conventions block, a link to docs/llms.txt, links to the four skill-reference recipes, and the new mapping page — making it the single router to every artifact.
• Skill references: api-cheatsheet gains a runnable "Borrowing external memory" recipe (TypedTensorView::from_slice wrapping a faer::Mat and an ndarray view, plus the TypedTensorViewMut::from_slice write-through variant); pitfalls gains per-origin traps (ndarray / nalgebra / PyTorch) and a redirect off the retired TypedStridedTensorView (Add TypedStridedTensorView for borrowed host strided tensor inputs #886). All mirrors (.claude, .kimi) kept byte-identical.
• Tooling: build_docs_site.sh republishes the skill references; check-docs-site.py adds reachability assertions (README links llms.txt and the skill; llms.txt links the README) plus a built-site skill-references check, with tests.

Verification

• python3 scripts/test-check-docs-site.py (new + existing checker tests)
• python3 scripts/check-doc-snippets.py --check, check-agent-skills.py (snippet/mirror sync)
• cargo test -p tenferro-tutorial-code (runnable interop snippet + all tutorial binaries)
• bash scripts/build_docs_site.sh /tmp/tenferro-site-1641 — full site build, step [8/9] runs the new checks and passes; acceptance items confirmed (llms.txt lists the mapping page + cheatsheet recipe + README; site serves skill-references/*.md and the rendered mapping page).
• bash scripts/check-pr-fast.sh --coverage-reviewed --test 'cargo test -p tenferro-tutorial-code' — passes.
• repository-rules-review — pass (LLM review skipped, local deterministic review).

Independent GPT review of the diff was incorporated (provider-feature additivity and reinterpret-vs-transpose wording corrected; reachability check hardened to parse link targets).

[github-actions]

Repository rules review

Repository rules review (659830c5cad3438b9fdc7705d26b75e97afcca14...44e4e8cb6f931b42817c8644839a81d8be233529)
Verdict: pass
Findings:
- [warn] llm-skipped (External LLM Review) <unknown>: External LLM review was skipped
  External LLM review is permanently disabled; deterministic checks only.