features: inspect_h5 keys, faster imports, archive UX

[youngdashu]

feat: inspect_h5 keys, faster imports, archive UX

What changed

• inspect_h5: In content mode, optional --name (comma-separated PDB keys) limits what is loaded and shown; order matches the list. Missing keys get a warning.
• CLI / imports: Lazy imports in command_parser, lazy embedder classes in EmbedderType, and deferred Embedding import in StructuresDataset so startup is lighter.
• Dataset path: If you pass a file as the dataset path, it must be .json.
• create_archive: Progress bar, results collected as workers finish, final zip name includes the dataset dir; no extra per-PDB files on disk beside the existing zip/tar flow.

Quick check

• inspect_h5 --mode content --name A,B and that --name without content errors.