Use released OSim tokenizer for SFT

[sunnweiwei]

This switches the SFT reproduction path away from code-level chat-template rewriting.

Changes:

• Reverts the Qwen3 chat-template override in sft/dataset.py.
• Makes run_sft.sh set actor_rollout_ref.model.tokenizer_path to the released OSim mid tokenizer by model size (cmu-lti/osim-4b-mid for Qwen3-4B, cmu-lti/osim-8b-mid for Qwen3-8B), while keeping ACTOR_TOKENIZER_PATH override support.
• Makes the SFT dataloader and SFT actor checkpoint metadata honor tokenizer_path, so saved checkpoints keep the selected tokenizer/chat template.

Validation:

• bash -n run_sft.sh
• python3 -m py_compile train_sft.py sft/dataset.py sft/sft_data.py sft/fsdp_worker.py sft/trainer.py
• Docker tokenizer check with Qwen/Qwen3-4B-Base + cmu-lti/osim-4b-mid: prefix-stable tokenization passes.
• Docker 8-GPU 1-step SFT smoke with Qwen/Qwen3-4B-Base + cmu-lti/osim-4b-mid: training and checkpoint save completed; saved tokenizer template matches cmu-lti/osim-4b-mid, with no last_query_index/<think>.