Handle bounding boxes in 2D ndimage warping

[sylvesterkaczmarek]

ndimage_warp assumes three spatial dimensions when reshaping map offsets, allocating an explicit output box and computing map/output offsets. With 2D images, explicit bounding boxes cause broadcasting errors or add an unwanted axis.

Make those operations dimension-aware. Add 2D regressions and 3D controls with translated/cropped boxes, unequal strides, output scaling and threaded warping.

Validation: python -m pytest -q tests/warp_test.py tests/map_utils_test.py tests/flow_utils_test.py passes all 32 tests. Five new 2D regression cases fail on unchanged upstream; all five 3D controls already pass. Syntax parsing and git diff --check pass.

[sylvesterkaczmarek]

@mjanusz This PR is current with main and has no public failing checks; only the Google import/review gate remains. It adds the 2D ndimage bounding-box handling with focused coverage. Could you review the current head when convenient?

[sylvesterkaczmarek]

@mjanusz Refreshed onto current main. The branch was 1 commit behind and is now 0 behind and mergeable; head bfa6f38fc9fcd60ac727843f4ce54df1a862cfc3 is GitHub Verified. The focused 2D/3D ndimage bounding-box regression class passes 10/10 selected tests. git diff --check is clean; the repository-wide Pyink reformat signal also reproduces on current main, so I left unrelated formatting unchanged. CLA is green and Copybara is waiting for internal safe-review approval. Could you review the refreshed head when convenient?