Randomize k-space split axis in Motion (#81)

[niravpatidar37]

Summary

Closes #81.

The Motion transform splits k-space into num_transforms + 1 segments and fills each from a differently moved copy of the image. Previously the split was hardcoded to the first spatial axis, even though the class docstring already documented it as being "along a random axis".

This PR randomizes the split axis, following the pattern already used by Ghosting:

• New axes=(0, 1, 2) constructor option. One axis is chosen at random per application.
• Pass axes=(0,) to keep the previous fixed-axis behavior.
• The chosen axis is recorded in the transform params ("axis"), so it is reproducible and works per-instance when the transform is applied to a batch.

Implementation notes

• The per-axis split is implemented with a vectorized per-element k-space mask (_segment_mask), so each element in a batch can use its own axis without a Python loop over elements.
• The degenerate-segment guard was generalized to _check_segment_sizes, which validates the size of whichever axis is actually used (error message wording unchanged).

Tests

• Updated the degenerate-segment test to use a volume that is too small on every axis, so it stays deterministic under a random axis, plus a pinned-axis variant.
• Added TestMotionAxes: axes validation, axis recorded in params, the split axis is genuinely randomized across applications, and pinned-axis reproducibility.
• tests/test_motion.py, tests/test_ghosting.py, tests/test_vectorization.py and tests/test_per_instance.py all pass; the vectorization suite confirms the per-element axis handling matches element-by-element application.