refactor: collapse eleven torchio wrappers into a table

[Hendrik-code]

Stacked on #61.

cpu/artifact.py and cpu/spatial.py held eleven functions between them — aug_motion, aug_ghosting, aug_spike, aug_bias_field, aug_blur, aug_noise, aug_swap, aug_flip, aug_affine, aug_elastic, aug_anisotropy — that were the same seventeen lines each, differing only in which tio.Random* they constructed:

def aug_ghosting(img, seg):
    if img.shape[0] == 2:
        subject = tio.RandomGhosting()(tio.Subject(...))   # <- only this line differs
        ...

They become two tables and one apply_tio call. −270 / +31.

It had already drifted

aug_anisotropy's single-channel branch passed axis=0 to tio.LabelMap:

tio.Subject(image=tio.ScalarImage(tensor=img), seg=tio.LabelMap(tensor=seg, axis=0))

Nothing else did that, and torchio stores it as unrecognised metadata rather than acting on it. It was most likely meant to be tio.RandomAnisotropy(axes=...). Collapsing the copies drops it — a behaviour change only in the sense that a typo stops being passed along.

Notes for review

• The tables hold factories, not instances, so each call still constructs a freshly seeded transform exactly as the functions did. There's a test for that.
• get_parameters in both classes had the same five lines of random_pick logic written out separately; that's torchio_ops.select now.
• apply_tio reads images out of the Subject by key rather than by attribute. torchio synthesises subject.image at runtime, which no type checker can see — the old code only passed mypy because those functions were unannotated and mypy skips unannotated bodies.

Testing

unit_tests/test_torchio_ops.py, 18 tests / 27 subtests, 8s.

It deliberately does not run all eleven transforms. apply_tio is the same code whichever transform it is handed, so running each would exercise torchio rather than this change — and I found out the hard way that RandomElasticDeformation(max_displacement=40) on a test-sized volume takes minutes, which would have put that straight into CI. Instead: every entry is checked for constructibility and freshness, the cheap ones run end to end, and apply_tio itself is tested against both input layouts (single-channel, and the two-channel "step 2" layout where channel 1 is an odd-disc segmentation) plus the property that LabelMap channels are never interpolated.

Full suite: 100 passed, 197 subtests. ruff and mypy clean.

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.