Census fold docs: reference collections, the TaxonMech sentence and the PDB term-space reason are wrong

[realmarcin]

From the round-3 review of #244.

Where: _fleet/README.md, the prefix_census.py comment, the build_subsets.py TERM_SPACE comment, and the PR body.

• The rule overclaims. "Every namespace named for a counted registry" would include GO_REF (3,031), GOC (40,653) and PO_REF (6). Those are reference or curator collections, which The vocabulary census misses most of the fleet's largest namespaces, not just TOGO #84 classes with literature. Folding GO_REF:0000058 into GO would fabricate a GO term. The rule has to name the exception.
• The TaxonMech sentence is wrong. "TaxonMech writes lowercase prefixes and the other Mechs upper-case forms": most of the lowercase spellings folded here come from other Mechs (kegg.compound, mediadive.medium, pdb.ligand, chembl, pubchem, uniprot.location), and TaxonMech also writes NCBITaxon:, GTDB: and DSM:.
• The PDB reason for TERM_SPACE is wrong. "Ids share values with the column's own ids" holds for RHEA-COMP (966 of 2,243 ids equal a reaction id) but not for PDB ligand codes. Those are 1–3 or 5 characters, and PDB entries are 4. Ligand codes keep their own key because they are a different kind of identifier and must not be shown as entries.

Fix: correct all three.