Review follow-ups: fail loudly on bad config, --json on usage errors, doctor hints - #50

Follow-ups from the review after #49. No version bump yet.

Fail loudly instead of falling back

• Config paths. load_config_from_file(path) raises when an explicit path doesn't exist. Before, it silently used the defaults, so the example scripts ran their ORB sections with UMA when started outside examples/. Called with no path, it still reads ./config.json if present and uses the defaults otherwise. The examples now resolve their configs and inputs from their own directory.
• Autobatcher bounds. memory_scaling_factor must be greater than 1, because otherwise the probe never grows. max_memory_padding must lie between 0.05 and 0.95. Both bounds are in the JSON Schema too.
• SevenNet D3.
  • Damping other than BJ is rejected; it used to be silently ignored.
  • D3 on an omol25_* modal logs the same double-counting warning as the other ωB97M-V models.
• Fairchem models need the omol task.
  • The CLI accepts gpuma's UMA list plus fairchem's *-omol models. r2scan, OC25 and ODAC25 are rejected.
  • The loaders check that the model has an omol task, which also covers model_path checkpoints.
• config --create writes the built-in defaults. It used to write the loaded ./config.json together with the CLI overrides. It now refuses to overwrite an existing file.

--json

• Argparse usage errors (exit 2) now print a summary too, with error.type UsageError, so --json always gives exactly one JSON object.
• The docs list checks, recommended_device, schema and installed.

doctor

• On an old driver with a CUDA 13 torch build, the reinstall hint adds --force-reinstall (pip) or --reinstall-package torch (uv). Without it, installing the same version is a no-op.
• For CUDA 12 builds, the hint says to update the driver.

Docs and housekeeping

• uv.lock is tracked, so uv sync --frozen reproduces the tested environment.
• docs/agents.md installs the tool with the cu126 index. This was tested on driver 535 and gives torch 2.13.0+cu126 with CUDA.
• Other docs fixes:
  • --model-type alone doesn't pick an ORB model name.
  • the --multiplicity help text.
  • the "Edit this page" link now points at master.
  • --python 3.12 in index.md.
  • the device default.
  • Morfeus wording.

Testing

• Full suite: 694 passed, 1 skipped.
• mkdocs build --strict passes.
• A single-atom input (Na⁺) on uma-s-1p2p1-ft-d4rev works in both paths, with matching energies.