bench: fix hardcoded per-dataset provenance metadata in loader.py

[davorrunje]

Context

benchmarks/datasets/loader.py:104 hardcodes metadata={"source": "zenodo:7968969", "license": "CC-BY-4.0"} for every dataset the loader returns. This mislabels datasets whose true provenance differs (e.g. adult is a UCI dataset, not that Zenodo record). It is currently inert — DatasetBundle.metadata is not consumed anywhere — but the field is wrong and would mislead anyone who starts using it.

Goal

Set correct per-dataset provenance metadata, or remove the misleading hardcoded value.

Where

• benchmarks/datasets/loader.py:104 (the hardcoded metadata=...).
• benchmarks/datasets/registry.py — where per-dataset info could be sourced.
• benchmarks/_common/bundle.py — DatasetBundle.metadata definition.

Acceptance criteria

• metadata["source"] reflects each dataset's true origin (per-dataset), or the hardcoded field is removed.
• No dataset is labeled with a provenance that isn't its own.

References

• PR feat(bench): HP-search sensitivity curves (spec + plan + implementation) #117 (noted during launcher/benchmark cleanup).