[Fix] Merge class lists when concatenating datasets with different types

[crawfordxx]

Motivation

Fixes #8890

When concatenating datasets with different class sets (e.g. CocoDataset + VOCDataset), ConcatDataset stored _metainfo as a list of dicts. This broke downstream evaluation code that expects metainfo to be a single dict with a classes key, causing IndexError in _det2json when cat_ids[label] was accessed with out-of-range labels.

Modification

Added a _merge_metainfo() method to ConcatDataset that:

• Merges class lists from all sub-datasets into a unified ordered set (preserving insertion order, deduplicating shared classes)
• Merges palette colours from each dataset, with a deterministic fallback for classes without a defined colour
• Returns a single merged metainfo dict instead of the list fallback, so evaluation works correctly with heterogeneous dataset concatenation

Also updated the full_init metainfo update logic to use isinstance check instead of the removed is_all_same flag.

BC-breaking (Yes/No)

No. When all datasets have the same classes, behaviour is unchanged. When datasets differ, the merged dict is a strict improvement over the previous list-of-dicts fallback which was unusable by downstream code.

Checklist

• Pre-commit hooks pass (pre-commit run --all-files)
• Unit test added (tests/test_datasets/test_dataset_wrappers.py)

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

---

majianhan seems not to be a GitHub user. You need a GitHub account to be able to sign the CLA. If you have already a GitHub account, please add the email address used for this commit to your account.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}