Semantic KITTI tutorial without explicit dataset export

[tejasstanley]

This PR updates the SemanticKITTI LiDAR segmentation tutorial to avoid exporting a converted train-ID dataset before evaluation.
The tutorial now keeps the original SemanticKITTI labels on disk and uses TorchLiDARSegmentationModel.eval() with ontology_translation and translation_direction="dataset_to_model" to translate labels in memory during evaluation.
Changes

• Removed the dataset.export(...) workflow from the SemanticKITTI tutorial.
• Removed the intermediate train-ID ontology file from the tutorial flow.
• Kept the dataset in the raw SemanticKITTI ontology.
• Added direct construction of the MMDetection3D output ontology from semantic-kitti.yaml.
• Added raw SemanticKITTI to MMDetection3D ontology translation.
• Added ignored raw classes for labels that map to unlabeled.
• Updated visualization/evaluation cells to use the direct raw-to-model translation.