Regroup every module into its layer directory

[isayev]

What

The six layers the dependency rules already enforced are now directories, so a module's location is its layer:

presentation/   cli/, auto3Dcli
entry/          auto3D, SPE, ASE/, tautomer
orchestration/  workflow, workflow_workers, chunk_manager, job_layout,
                processors, pipeline/
engines/        models/, model_factory, isomers/, batch_opt/
domain/         ranking, filtering, embedding, clash_relief, id_mapping
foundation/     config, constants, exceptions, registry, results,
                torch_config, utils/

79 modules moved, 175 files rewritten. Layer packages are namespacing only and re-export nothing, per the barrel rule.

What this buys

The layer map in tests/test_layer_boundaries.py was a per-module prefix list that had to be edited on every move, and could go stale in both directions: a module in no layer (caught from the start) and a layer prefix naming no module (uncaught until #169 added a test for it — after #168 left isomer_engine behind).

Both failures are now inexpressible. The map is derived from directory names, and the only thing still declared is the order of the layers, which is the one fact a directory name cannot carry. The staleness test shrank to "does every layer directory exist" — a layer can no longer be partly stale.

The top-level API is unaffected

from Auto3D import main, smiles2mols, Auto3DOptions, calc_spe, calc_thermo, opt_geometry, create_model, ... all still work. The package root re-exports them lazily through _LAZY_API and its contents did not change.

That is what turns this from "every user's imports break" into "code reaching into submodules breaks" — and it rests on one dictionary, exercised by test_public_api.py and by the api.rst path-resolution test rather than merely eyeballed.

Breaking for module paths, with a CHANGELOG table. The exception classes are the ones most likely to appear downstream: all eight moved to Auto3D.foundation.exceptions. No shims — a shim would give every moved name two supported spellings, which is the rule the barrels already state.

Five kinds of reference a dotted-path rewrite does not find

All five hit here, and they're the transferable part of this PR:

kind	example	caught by
submodule imported by name	from Auto3D import model_factory — the text Auto3D.model_factory never appears (10 sites, 5 files)	tests
filesystem fragments	_submodules("utils"), SRC_ROOT / "isomers" / ..., a checkpoint path read off disk	tests
assertion literals	["Auto3D/isomers/base.py"] as an expected value	tests
[project.scripts]	the auto3d console-script entry point	nothing
[tool.setuptools.package-data]	models/*.pt — the wheel would have shipped without the ANI2xt checkpoint	nothing

The last two matter most: tests pass from the source tree regardless, so a green suite says nothing about either. I found them by reading pyproject.toml, not by running anything.

Documentation

api.rst and the 25 checked-in autosummary stubs were renamed and rewritten — Sphinx keys them by dotted path, so titles and underlines moved too. The current-usage docs were rewritten.

The two migration guides were deliberately not. Both are before/after records — migration.rst's "before" is v2.x code — so rewriting their "after" halves would make each describe a layout that did not exist when the change it documents was made. Each now carries a note saying its paths are era-specific and pointing at the CHANGELOG table.

Verification

• 1771 passed, 1 skipped, 74 deselected — the same count as before the move, randomized order.
• mypy: 68 errors, unchanged; 85 files checked instead of 79, which is the six new layer __init__.py files.
• test_no_module_imports_a_higher_layer still passes, so the layering that was enforced against the old declared map holds against the derived one.