feat: add tensor-field constructor for EllipseCloud

[t-uda]

Motivation

DTI-style workflows currently require manual slicing, masking, reshaping, and placeholder neighborhood handling. A first-party helper would remove boilerplate for diffusion tensors, structure tensors, and stress/moment tensors.

Scope

• Provide a constructor that accepts tensor-field data plus an optional mask.
• Define a clear path from tensor grid coordinates to ellipse centers and covariances.

Acceptance Criteria

• A masked tensor field can be converted into an EllipseCloud without manual coefficient assembly.
• The API documents how centers are generated and how tensor submatrices are chosen for 2D workflows.
• Add at least one regression test based on a small synthetic tensor field.

Open Design Questions

• How affine should affect centers.
• Whether the helper should target 2D slices first or support general (..., d, d) tensor fields from the start.