BUG: Pin AdaptiveNonLocalMeansDenoisingImageFilterTest1 to single work unit

itkAdaptiveNonLocalMeansDenoisingImageFilter::ThreadedGenerateData scatter-
writes m_RicianBiasImage at neighborhood-offset indices (centerIndex +
patchOffset) that cross work-unit boundaries, so multiple work units perform
unsynchronized SetPixel on the same shared pixels. The data race is benign at
low thread contention but, under CPU oversubscription (ctest -j8), preemption
widens the window and yields nondeterministic output: the test fails the
baseline comparison intermittently (ImageError != 0), yet passes in isolation.

Verified: with the default work-unit count the test fails 50-100% of -j8 runs;
pinned to a single work unit it passes 100% (6/6) under the same load and
matches the baseline (which reflects the race-free result).

Pin the test filter to one work unit for determinism until the filter's scatter
is made race-free (the proper long-term fix belongs in the filter, not the test;
tracked in #6419).

Author: hjmjohnson

Modules/Filtering/AdaptiveDenoising/test/itkAdaptiveNonLocalMeansDenoisingImageFilterTest.cxx

@@ -115,6 +115,10 @@ itkAdaptiveNonLocalMeansDenoisingImageFilterTest(int argc, char * argv[])
   ITK_EXERCISE_BASIC_OBJECT_METHODS(filter, AdaptiveNonLocalMeansDenoisingImageFilter, NonLocalPatchBasedImageFilter);
 
   filter->SetInput(reader->GetOutput());
+
+  // Single-threaded: m_RicianBiasImage data race across work-unit boundaries; see #6419.
+  filter->SetNumberOfWorkUnits(1);
+
   filter->SetUseRicianNoiseModel(false);
   ITK_TEST_SET_GET_VALUE(false, filter->GetUseRicianNoiseModel());