dataset-names adopts the source toolkit (spec 004) #5

Follow-up of spec 004 (#4): dataset-names replaces its own copies of the generic source code with the benchcommons-tools-dataset package. The dataset-names side (spec, plan, code) is written in dataset-names; this issue tracks it from tools-dataset.

Contract to adopt: docs/specs/004-source-toolkit/spec.md in this repo, at tag v0.1.0.

• Tag v0.1.0 of tools-dataset exists (created after spec 004 merges)

• dataset-names spec for the adoption, linking to spec 004 here instead of redefining it

• Dependency declared as in ADR 0004 of this repo ([tool.uv.sources] git source pinned to the tag)

• src/namescommons/cache.py replaced by datasettool.sources.cache (User-Agent of dataset-names passed explicitly)

• src/namescommons/spreadsheet.py replaced by datasettool.sources.spreadsheet

• Generic part of src/namescommons/adapters/__init__.py replaced by datasettool.sources.adapters; RawRecord and RawLink stay in dataset-names; check_adapter called with optional=("links",)

• src/namescommons/sources.py rebuilt on datasettool.sources.registry with its own fields (country, count_type, primary, suppression, uniform_case), its accepted licences (ADR 0001 and 0004 of dataset-names) and the one-primary check

• src/namescommons/manifest.py uses datasettool.files for checksums and CSV row counts

• Ported tests removed from dataset-names only after the equivalents pass here (list them before deleting)

• Do not assert on the order of lines inside a multi-error registry message: the toolkit checks retrieved and adapter before the dataset fields (country, count_type, primary, suppression, uniform_case), while dataset-names checked them in between (plan 2.4 of spec 004); the set of messages is the same

• The cache now refuses source ids outside [a-z0-9]+(-[a-z0-9]+)* and unsafe file names (ValueError), as dataset-names' registry already does

Done in BenchCommons/dataset-names#41 (spec 007, ADR 0013): every item above is implemented and verified there (identical outputs on the fixtures and on the full real cache). Follow-ups: #9 (the cache refuses symlinked entries, found during the adoption) and BenchCommons/dataset-names#40 (message language).