feat(provenance): add PROV-O RDF export for BrainKB ingestion

[dhritimandas]

Summary

What: This adds a command. The command is nobrainer provenance export.
It reads a saved model bundle. It emits a PROV-O graph in Turtle or JSON-LD.
Why: No structured provenance graph existed for a training run.
croissant.json records model and dataset facts, but nothing links them as
an activity. BrainKB needs that graph to ingest nobrainer runs.

The graph

The run is a prov:Activity. It prov:used a dataset entity and
prov:generated a model entity. Both link to one or more
prov:wasAssociatedWith agents (always the nobrainer/pytorch software
agents; optionally a caller-supplied human/org agent via --agent, never
invented by this module on its own).

Determinism

Every IRI is content-derived: a file/model's own sha256, or a stable hash
of a canonical-JSON payload for the run/dataset/agent nodes. No timestamps,
no random UUIDs, no blank nodes anywhere. Re-exporting the same input from
a different location on disk produces a byte-identical graph -- verified
directly, not just asserted (test_run_iri_unchanged_after_directory_move,
test_repeated_export_is_byte_identical).

DOMAIN provenance only -- enforced, not just documented

This module emits facts about what nobrainer itself did. It does not
emit BrainKB ingestion-activity triples (e.g. "this graph was loaded into
BrainKB at time T"). Named-graph registration and POSTing this output to
BrainKB's own ingestion API are separate steps the caller performs against
BrainKB's API. A runtime check (_assert_domain_only, called on every
build_graph()) rejects: a second prov:Activity, any agent not
associated with the run, forbidden PROV predicates
(wasInformedBy/wasStartedBy/wasEndedBy/qualifiedAssociation/
hadPlan), and any term containing "brainkb". A positive-control test
injects a second activity and confirms the check actually fires.

Inputs

• --bundle DIR (required): a Segmentation.save() directory
  (model.pth + croissant.json). Read with plain json.load -- not via
  nobrainer.processing.croissant, since importing that module pulls in
  torch through processing/__init__.py (verified). This module is
  stdlib + rdflib only.
• --dataspec PATH.json (optional): a DataSpec manifest
  (nobrainer/data/spec.py). When given, dataset provenance includes both
  image and label checksums. When omitted, falls back to croissant's
  source_datasets, which only ever records image paths.
• --agent, --base-iri, --strict, --format {turtle,json-ld}.

Edge cases handled explicitly, not silently

• A non-finite final_loss/best_loss is omitted (nb:lossStatus "non-finite") rather than serialized: rdflib emits a bare NaN in
  JSON-LD, which is not valid JSON (verified).
• A model_args/optimizer.args value that landed in croissant.json as a
  Python repr() memory address (write_model_croissant uses
  default=str) is elided rather than hashed -- it would otherwise make
  run IRIs nondeterministic across machines.
• Empty block_shape/None fields emit no triple (absence, not "").
• Missing checksums fall back to on-disk recomputation, then to a visibly
  unaddressed x-nosha/ IRI, recorded via nb:checksumStatus.
• write_model_croissant and write_checkpoint_croissant disagree on
  model_architecture vocabulary (registry name vs. torch class name);
  the disagreement is recorded (nb:architectureVocabulary), not guessed.

The one new dependency

rdflib was already present transitively (via mlcroissant) but
undeclared. Now [project.optional-dependencies] provenance = ["rdflib >= 7.0"], added to all.

Known gap, flagged rather than papered over

training_date in croissant.json is the model-save timestamp, not a run
start/end. It is emitted as prov:generatedAtTime on the model (where
it's literally true) and as a neutral nb:metadataRecordedAt on the run
-- never as prov:endedAtTime, which would misrepresent it. The graph
genuinely cannot express training duration until write_model_croissant
records real start/end times.

Test plan

• uv run pytest nobrainer/tests/unit/test_rdf_export.py -q -- 52
  passed (hashing primitives, graph shape, --agent, --dataspec,
  edge cases, the DOMAIN-only boundary with a positive control,
  turtle/JSON-LD equivalence, determinism, no-torch-import isolation,
  CLI subprocess tests).
• uv run pytest nobrainer/tests/unit/ -m "not gpu" -q -- 446 passed,
  no regressions.
• pre-commit run --all-files clean.
• Real end-to-end run against an actual bundle: Turtle and JSON-LD both
  re-parse cleanly with a fresh rdflib.Graph().parse() and produce
  the same triple count (81 == 81).
• CI