Streamline training and export tutorials

[yutong-xiang-97]

What has changed and why?

• Streamline the LTDETRv2 and EoMT training tutorials around a single train → best-checkpoint → predict workflow.
• Add complete training, validation, checkpoint reload, and prediction sections for EoMT instance, semantic, and panoptic segmentation.
• Add the small dataset preparation steps required for the EoMT notebooks to run end to end in Google Colab.
• Remove the duplicate pretrained-model prediction walkthroughs from the training tutorials.
• Pin the tested TensorRT 10.13.3.9 package in the export tutorials and avoid redundant ONNX verification in the EoMT instance-segmentation export path.

This keeps the tutorials focused on training a model and immediately using the exported checkpoint, while making the Colab workflows reproducible.

How has it been tested?

• Ran all the notebooks end to end on a Google Colab T4.

Did you update CHANGELOG.md?

• Yes
• Not needed (internal change)

Did you update the documentation?

• Yes
• Not needed (internal change without effects for user)

[yutong-xiang-97]

/review

[liopeer]

Small potential issue, otherwise lgtm!