Add pipeline transform order validation

[crawfordxx]

Summary

Fixes #6106

As reported in the issue, the ordering of data augmentation transforms in training pipelines is error-prone and can silently produce incorrect results (e.g., zero evaluation metrics). This PR adds a validate_pipeline_order() utility that:

• Warns when transforms appear in a suspicious order (e.g., Normalize before RandomFlip, Pad before Resize)
• Uses both pair-wise rules and category-based ordering checks
• Is called automatically in BaseDetDataset.__init__ before the dataset is built
• Only emits UserWarning messages — never raises exceptions — so existing configs continue to work

Changes

• mmdet/datasets/utils.py: Added validate_pipeline_order() with transform category definitions and pair-wise ordering rules
• mmdet/datasets/base_det_dataset.py: Calls validate_pipeline_order() during dataset initialization
• mmdet/datasets/__init__.py: Exports the new function
• tests/test_datasets/test_utils.py: Unit tests covering correct pipelines, the exact problematic pipeline from Training pipeline transformation order? #6106, and various misordering scenarios

Example warning

When using the problematic pipeline from #6106 (Normalize before Pad):

UserWarning: In the data pipeline, 'Pad' (index 5) appears after 'Normalize' 
(index 4), which is likely incorrect. Normalize should be applied after Pad so 
that padded values are also normalized. Please verify the transform order in 
your config.

Test plan

• Correct standard pipeline produces no warnings
• Issue Training pipeline transformation order? #6106 pipeline (Normalize before Pad) triggers warning
• Normalize before RandomFlip triggers warning
• Pad before Resize triggers warning
• Empty pipeline does not crash
• Unknown custom transforms are silently skipped