Remove obsolete code and files and simplify developer tooling

[kierandidi]

The benchmark wrapper could report success after pytest failed, and revision comparison switched/stashed the caller's checkout while ignoring failures. Build each revision in an isolated checkout/environment, preserve local work, and retain benchmark results, logs and source/environment provenance.

Remove the top-level notebooks folder, retaining the useful model-input example under docs with all consumers updated. Retire obsolete developer tools, eight unused native headers, a dead component plot, redundant profiler indirection and a forwarding-only test helper. Move test dependencies out of runtime requirements, stop eager vendored-test imports, fix vals(), remove stale configuration and MANIFEST.in, and index scientific generators/derivations. Whole-tree analysis covered 2,281 tracked files and the 87-translation-unit/137-header native include graph; public compatibility APIs and scientific assets remain intact.

Validation: 132 packaging tests; 25 public-API/mapping tests; two fresh isolated native comparison builds with ten CPU benchmarks each; fresh runtime-only scoring, gradients, repacking and minimization; model-input notebook CPU execution; full strict Sphinx build. After the final native cleanup, rebuilt the affected CPU library and bonded-atom extension, then passed 89 focused IO/scoring/gradient/sampler/API tests. Formatting/lint/diff checks pass. Removing MANIFEST.in preserved all other real sdist contents.

CUDA rebuilding/testing for the final cleanup revision remains pending CI. Normal checks and required reviews remain the merge gate; no performance claim is made by this cleanup PR.