Fix parser crashes, query generation bugs and add_path regression

[srmnitc]

Fixes from an audit of tools4RDF 0.3.8. There is one commit per fix; each bug fix adds a regression test to tests/test_regressions.py, and I checked that each test fails without its fix. All 181 tests pass (146 existing + 35 new).

Crashes and silently wrong results

• Union members leaked across parses. unravel_relation had a mutable default list, so members of every owl:unionOf accumulated for the whole process. After one ontology with unions was loaded, later parses got wrong subclasses or crashed with KeyError.
• PMDco could not be loaded (full, base or minimal). parse_subclasses and parse_named_individuals raised KeyError on nested class expressions and on parents from ontologies that aren't loaded. Those are now skipped.
• add_path did nothing. Since the lazy-graph restructure it added the literal triple to the ontology graph, which the parser ignores, so atomRDF's rdfs:label paths were dead. It now records the path as rdfs:domain/rdfs:range on the property. That means it applies to subclasses and survives rebuilding and +. Unknown properties raise ValueError.
• Multi-source queries lost their FILTER. _add_filters reset the destination terms after the first source, so every later query ran unfiltered. _modify_destinations also wrote stepped-query parents onto the shared terms in terms, and these stayed behind if a query raised. Destinations are now copied instead.
• create_query(op, [d1, d2]) raised TypeError. _is_already_in_destinations was called with *destinations.
• Wrong IRIs from namespace lookup. _lookup_namespace returned the first matching namespace rather than the longest, so with base: <http://example.org/> and ex: <http://example.org/o#> bound, ex:A became base:A.
• Colliding variables joined unrelated values. Terms with the same local name (ex:value, y:value) shared one variable. Colliding variables are now prefixed with the namespace; all other names are unchanged.

Generated SPARQL

• A FILTER using xsd: was emitted without PREFIX xsd:. rdflib tolerates this; spec-strict engines (e.g. Oxigraph) reject it.
• Filter values were pasted into strings. Now:
  • quotes are escaped
  • booleans use "true", not "True"
  • an rdfs:Literal range no longer becomes xsd:Literal, which matched nothing
  • a data property without a range no longer raises IndexError
• Intermediate classes with hyphens (six OCDO classes, e.g. pldo:Twin-likeStackingFault) produced invalid variables.
• The start class for object-property sources was picked from a set, so it changed with PYTHONHASHSEED. PREFIX lines are now sorted.

API

• OntoTerm:
  • label was always None; labels are now also read from rdfs:label
  • terms were unhashable
  • comparing with a non-term raised an error
  • and_/or_ returned None
• Named individuals get node_type="named_individual". create_query raises for sources that are not classes or object properties; previously it silently returned []. A term that is both a class and an individual keeps its class entry.
• add_term expands prefixed names in dm/rn; dm=["ex:A"] used to create a disconnected :ex:A node.
• copy.deepcopy(onto.terms) no longer recurses infinitely.
• visualize_graph works with its default style, and no longer title-cases literal values.
• owl:Thing is no longer listed as its own subclass.
• Docstrings of query() and add_namespace() now match their behaviour.

Packaging and CI

• Removed dict | dict (the package declares Python >= 3.8).
• Bundled-ontology path built with os.path, so it works on Windows.
• Removed the unused numpy dependency.
• The Codecov slug pointed at pyscal/atomRDF; now OCDO/tools4RDF. actions/checkout@v4.
• Removed an incorrect author email in pyproject.

Behaviour changes

• query(url, ...) sends the query directly to the endpoint (SPARQLStore, form POST) instead of wrapping it in SERVICE over an empty local graph. LIMIT and FILTER now run on the server; previously rdflib fetched every row and applied LIMIT locally. create_query(remote_source=...) still produces the SERVICE form. A string that is not a URL raises ValueError.
• Result column names change only when two local names collide (e.g. ex_valuevalue, y_valuevalue).
• Using a named individual as a source raises ValueError instead of returning [].

Not in this PR

• The bundled data/cmso.owl is out of date: 12 upstream terms are missing, and 5 terms have since been removed upstream. Updating it changes which terms are available, so it should be a separate decision.
• examples/03_querying_materials_ontologies.ipynb already fails on main. It needs pmd_2_ontology.ttl, which is not in the repository, and it reads dataset/triples, while the file is examples/dataset/ocdo_triples.
• & and | still clear the conditions of their operands.

Verification

Besides the test suite:

• PMDco full, base and minimal load.
• atomRDF's six-ontology setup, built from local copies of the OCDO ontologies, reaches rdfs:label from asmo:CalculatedProperty and its subclasses.
• A remote query against https://matkg.pyscal.org/sparql returns the expected rows.
• Example notebooks 01, 02 and pizza run.