Make GPU NDT covariance updates atomic under contention

[fredfrog78]

Summary

• harden covarianceHitNdt so only one thread owns each voxel update, including an ungrouped-sample fallback path
• pass a grouped_samples kernel argument from GpuNdtMap so the kernel can choose fast grouped traversal vs safe ungrouped ownership scan
• add regression coverage for ungrouped contention (Ndt.HitCountUngrouped)
• include portability/build fixes encountered while completing test execution in this environment

Files changed

• ohmgpu/gpu/CovarianceHitNdt.cl
• ohmgpu/GpuNdtMap.cpp
• tests/ohmtestgpu/GpuNdtTests.cpp
• cmake/Findglm.cmake
• ohm/CovarianceVoxel.h
• ohmgpu/private/GpuMapDetail.cpp

Test results run locally (Apple M3 Pro, OpenCL 1.2)

• ✅ cmake -S . -B build3 -DOHM_FEATURE_TEST=ON -DOHM_SYSTEM_GTEST=ON
• ✅ cmake --build build3 --target ohmtestocl -j8
• ✅ ./build3/bin/ohmtestocl '--gtest_filter=Ndt.*'
• ❌ full ./build3/bin/ohmtestocl has 2 failures:
  • Tsdf.Basic
  • Tsdf.Truncation

These TSDF failures are due to the local OpenCL runtime limitation (cl2Metal / OpenCL 1.2 lacks required 64-bit atomics), not this NDT change.

Maintainer request before merge

Please run maintainer validation on a supported GPU/runtime (e.g. CUDA target and/or OpenCL 2.0+ environment) before merge, including:

• ohmtestocl (or equivalent supported OpenCL target)
• ohmtestcuda where available

[fredfrog78]

Closing for now per review request: will resubmit after dependencies are fixed and tests pass end-to-end.

[fredfrog78]

Maintainer note:
Local full ohmtestocl run is blocked by platform-specific TSDF OpenCL 1.2 limitations (Tsdf.Basic and Tsdf.Truncation).

Please run maintainer-side GPU/runtime validation before merge on supported targets (e.g. CUDA and/or OpenCL 2.0+ environments).