feat: add strict GraphTokenizer paper reproduction workflow

[baihchou8787]

Summary

• add a strict GraphTokenizer paper protocol pinned to official source revision 98343a6b025a48fbb6859cd812a12b81ec3ac3cc
• match the paper data, optimization, augmentation, 100-serialization, test-time averaging, and five-run sample-standard-deviation settings
• report QM9 HOMO MAE after inverse scaling in raw eV, MolHIV ROC-AUC from averaged positive-class probabilities, and Peptides-struct mean MAE after per-target inverse scaling
• require the pinned official GTE encoder checkpoint and record reproducibility metadata
• add regression coverage for serializers, tokenizers, pooled augmentation paths, losses, normalization, aggregation, schedules, and strict argument validation

Validation

• TL_BACKEND=torch python -m pytest -q tests/models/test_graph_tokenizer_paper_protocol.py tests/models/test_graph_transformer.py tests/transforms/test_graph_tokenizer.py
• 188 passed, 2 skipped
• all six BERT/GTE × QM9/OGBG-molhiv/Peptides-struct strict preflight checks passed in the pinned paper runtime
• pinned GTE checkpoint SHA-256 verified

Local experiment logs, caches, and checkpoints are excluded from this PR.