Position embeddings allocated but never used - model is bag-of-tokens

[consigcody94]

Found via code audit. transformer.py:52. x = tok_emb directly without adding position_embedding. Model cannot learn sequence order. All training produces garbage.