Separate unit tests from integration tests

Running pytest with no arguments currently downloads a tarball from ECMWF's S3, unpacks it, computes standardization stats, builds graphs, and runs a full Lightning training loop. That's a lot to ask for a quick local check.

Beyond the slow feedback loop, mixing test types in a flat directory makes it harder to understand what each test is actually verifying. A reader can't tell at a glance whether a test is checking isolated logic or exercising the full stack, which makes the suite harder to maintain and extend.

Most tests call init_datastore_example() from conftest.py, which triggers the S3 download. Only a handful are actually fast and isolated:

• Unit tests (no I/O): test_imports.py, test_config.py, test_cli.py, test_time_slicing.py, and the test_all_gather_cat_* functions in test_training.py
• Integration tests (need real data): test_datastores.py, test_graph_creation.py, test_training.py::test_training, test_datasets.py, test_plotting.py, test_plot_graph.py

Options

1. Register a marker in pyproject.toml:

[tool.pytest.ini_options]
markers = [
    "integration: tests that require network access or real data (deselect with '-m not integration')",
]

2. Split into tests/unit/ and tests/integration/, moving conftest.py (and the S3 fixture) into the integration folder. Cleaner long-term, makes the distinction visible in the file tree, and prevents the fixture from being accidentally used in unit tests.

Bonus

test_clamping.py could become a true unit test by swapping the real MDPDatastore for the DummyDatastore that already exists in tests/dummy_datastore.py. Same logic, no S3 dependency. test_time_slicing.py already does this well and is a good model to follow.