Clarify evaluation loaders and unify composite model loading

[xieofxie]

Context

_select_model_loader() in src/winml/modelkit/eval/evaluate.py selects several distinct loading contracts, but _ModelLoaderKind does not document what each kind means or who owns model/session construction.

Mask generation currently selects EVALUATOR_MANAGED: WinMLMaskGenerationEvaluator constructs its encoder and decoder ONNX Runtime sessions directly. This bypasses WinMLAutoModel because the composite model type is not registered. That makes the loading design task-specific and difficult to reuse for future evaluators backed by composite models.

Work

• Document every _ModelLoaderKind value, including its input form, loading owner, returned model shape, build-pipeline behavior, and intended evaluation mode.
• Investigate and design an architecture-agnostic composite-model loading path through WinMLCompositeModel / WinMLAutoModel.
• Migrate mask-generation evaluation to the shared composite-model path if feasible.
• Ensure future composite evaluators can reuse the same path without adding task- or architecture-specific branching to _select_model_loader().
• Add pytest coverage for loader selection and composite-model loading behavior.

Acceptance criteria

• Loader-kind semantics and precedence are clear at their declaration.
• Mask-generation session ownership is either moved into a reusable composite model abstraction or the remaining blocker is explicitly documented.
• The design does not hardcode model architecture, graph node, tensor, or layer names in shared loading code.
• Existing ONNX, ONNX-to-HF comparison, ONNX-to-ONNX comparison, and GenAI loading behavior remains unchanged.