feat(yolo26-sem): binary (nc==1) post-process + register fine-tuned variants in core registry

[leeclemnet]

What does this PR do?

Makes YOLO26 semantic-segmentation models (served via inference_models) work end-to-end on the inference server, and tightens the semantic-seg class-name contract. Four changes:

1. Core registry (inference/models/utils.py): ROBOFLOW_MODEL_TYPES registered only the family key ("semantic-segmentation", "yolo26") — which pretrained/public models match (modelArchitecture = "yolo26"), but fine-tuned versioned models report their ORT modelType ("yolo26n-sem", …) with no variant→family normalization, raising ModelNotRecognisedError. Now registers the family alias plus every trained variant, mirroring the YOLOLite registration.
2. Binary (nc==1) post-process: a single-foreground-class model trains as Ultralytics' native binary head → one output channel. softmax over one channel is degenerate (≡1.0), so the single-channel case uses the sigmoid foreground probability; sub-threshold pixels collapse to background via the existing threshold step. Shared by ONNX/TorchScript/TRT; multi-channel (DeepLab + multiclass) is untouched (only triggers on shape[0] == 1).
3. Load-time validation: a shared validate_class_names requires a semantic-seg package to declare background + ≥1 foreground class (>= 2 names), raising CorruptedModelPackageError instead of failing silently. resolve_background_class_id is now a pure lookup that assumes a validated package.
4. Refactor: moved the semantic-seg class-name helpers out of the generic model_packages.py into common/roboflow/semantic_segmentation.py (shared by the DeepLabV3+ and YOLO26-sem backends).

Related Issue(s): Builds on #2372 (YOLO26-sem ONNX/TorchScript/TRT backends). Pairs with roboflow-train #807.

Type of Change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Refactoring (no functional changes)

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:
pytest over the semantic post-process, model_packages, and the new semantic_segmentation tests → 53 passed. Binary tests verified meaningful by direct invocation (without the branch a sub-threshold pixel wrongly becomes foreground [1]; with it, background [0]). New tests: test_validate_class_names_* (valid binary/multiclass, missing background, missing foreground). Registry resolution of ("semantic-segmentation", "yolo26n-sem") is covered by CI's full-deps import. E2E on staging (fine-tuned single-class model, paired with roboflow-train #807) pending.

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable) — n/a

Additional Context

Multi-channel/DeepLab behavior is unaffected throughout (binary branch only triggers on a single output channel). post_process_semantic_segmentation_logits intentionally stays in post_processing.py (it's a post-processor); only the class-name validation/resolution helpers moved.