Radiology inference breaks on MONAI 1.6 (Spacingd and NibabelWriter regressions upstream)

Problem

requirements.txt pins only monai>=1.4.0, so a fresh install picks up MONAI 1.6.x. Two MONAI core regressions in 1.6 affect MONAI Label:

Spacingd (Spacingd crashes when the data dict has a non-dict key starting with '<key>_' (regression in 1.6.0) MONAI#9165, fix proposed in Fix Spacingd updating non-metadata entries that share the key prefix MONAI#9167). Spacingd now calls .update() on every "<key>_*" entry. BasicInferTask always adds the string data["image_path"] (monailabel/tasks/infer/basic_infer.py), and the radiology sample app uses Spacingd("image", ...) in many infer and train pipelines. As a result, NIfTI inference raises AttributeError: 'str' object has no attribute 'update'. This is what made tests/unit/scribbles/test_transforms_infer.py::test_histogram_graphcut_inferer_* fail.
NibabelWriter with 2D arrays (NibabelWriter fails on 2D channel-less arrays with no target affine (regression in 1.6.0) MONAI#9166). Writing a 2D channel-less array with NibabelWriter and no target affine raises TypeError: '>=' not supported between 'NoneType' and 'int'. This affects monailabel.transform.writer.write_nifti.

Both work on MONAI 1.5.1.

Item 2 has a behaviour-preserving workaround in write_nifti: pass the input spatial shape explicitly when there is no target affine. Output is identical on 1.5.1 and 1.6.1.
Item 1 must be fixed in MONAI (Fix Spacingd updating non-metadata entries that share the key prefix MONAI#9167). Until a fixed MONAI is released, the two graphcut tests are marked non-strict xfail on MONAI >= 1.6. Once MONAI releases the fix, remove the xfail and consider raising the minimum MONAI version.