Add SYCL support for t/geometry/t/pipelines kernels, Transform ops, and NNS CPU fallback

[Copilot]

Port missing CUDA kernels to SYCL and fix all "Unimplemented device" errors for SYCL tensors in the t/pipelines and t/geometry subsystems.

Type

• New feature (non-breaking change which adds functionality). Resolves #

Motivation and Context

SYCL:0 devices hit utility::LogError("Unimplemented device") in several hot paths:

• Transform{Points,Normals} / Rotate{Points,Normals} — no SYCL dispatch in Transform.cpp
• EstimateNormals, EstimateColorGradients, RemoveRadiusOutliers, RemoveStatisticalOutliers, ComputeBoundaryPoints — NearestNeighborSearch asserts non-SYCL
• All t/pipelines/kernel functions (ComputeFPFHFeature, ComputePosePoint*, FillInLinearSystem, RGBDOdometry, TransformationConverter) — no SYCL kernels existed

Checklist:

• I have run python util/check_style.py --apply to apply Open3D code style
  to my code.
• This PR changes Open3D behavior or adds new functionality.
  • Both C++ (Doxygen) and Python (Sphinx / Google style) documentation is
    updated accordingly.
  • I have added or updated C++ and / or Python unit tests OR included test
    results (e.g. screenshots or numbers) here.
• I will follow up and update the code if CI fails.
• For fork PRs, I have selected Allow edits from maintainers.

Description

New SYCL pipeline kernels (t/pipelines/kernel/)

Five new *SYCL.cpp files implement SYCL equivalents of every CUDA kernel:

File	Functions
FeatureSYCL.cpp	ComputeFPFHFeatureSYCL
FillInLinearSystemSYCL.cpp	FillInRigidAlignmentTermSYCL, FillInSLACAlignmentTermSYCL, FillInSLACRegularizerTermSYCL
RegistrationSYCL.cpp	ComputePosePointToPointSYCL, ComputePosePointToPlaneSYCL, ComputeColoredICPResidualAndGradientSYCL, ComputePointToPlaneDistancesSYCL
RGBDOdometrySYCL.cpp	All four RGBD odometry pose solvers
TransformationConverterSYCL.cpp	Pose↔transformation converters

Kernels use sycl::nd_range + sycl::reduce_over_group for reduction paths (matching CUDA's cub::BlockReduce) instead of plain global atomics, and sycl::local_accessor for shared local memory.

New SYCL geometry kernel (t/geometry/kernel/)

TransformSYCL.cpp — SYCL implementations of TransformPoints, TransformNormals, RotatePoints, RotateNormals via sycl::queue::parallel_for. Reuses the existing TransformImpl.h per-element kernels, guarded by a new OPEN3D_SKIP_TRANSFORM_MAIN macro (mirrors OPEN3D_SKIP_FPFH_MAIN in FeatureImpl.h) to avoid duplicate symbol errors.

Transform.cpp and Transform.h updated with IsSYCL() dispatch branches.

NNS CPU fallback for SYCL devices (t/geometry/PointCloud.cpp, t/pipelines/registration/)

NearestNeighborSearch explicitly rejects SYCL tensors. Every NNS call site now transparently redirects to CPU when device.IsSYCL():

const core::Device nns_device = device.IsSYCL() ? core::Device("CPU:0") : device;
const core::Tensor positions_nns = GetPointPositions().To(nns_device);
core::nns::NearestNeighborSearch nns(positions_nns);
// ... run NNS ...
indices = indices.To(device);  // bring results back for SYCL kernels

Affected: EstimateNormals (Hybrid/KNN/Radius), EstimateColorGradients (Hybrid/KNN/Radius), RemoveRadiusOutliers, RemoveStatisticalOutliers, ComputeBoundaryPoints, ComputeFPFHFeature, CorrespondencesFromFeatures, EvaluateRegistration, MultiScaleICP, GetInformationMatrix.

Test coverage

All t/pipelines tests (Registration, Feature, TransformationEstimation, RGBDOdometry, TransformationConverter, SLAC) updated to use PermuteDevicesWithSYCL so SYCL:0 is exercised in CI.

---

💡 You can make Copilot smarter by setting up custom instructions, customizing its development environment and configuring Model Context Protocol (MCP) servers. Learn more Copilot coding agent tips in the docs.

[update-docs]

Thanks for submitting this pull request! The maintainers of this repository would appreciate if you could update the CHANGELOG.md based on your changes.