Fix tokenizer_config: use minimal clean T5 config (remove backend=tokenizers)

Fix: remove extra_special_tokens (incompatible list format) - extra_ids=100 handles this

Fix tokenizer_config: convert extra_special_tokens from list to dict for transformers compat

Update to v4: trained on 6.2K samples, composite 0.6047, 8/10 Human-Written verdicts

Upload tokenizer

Training in progress, step 500