Decouple Megatron paths and centralize chat-template defaults

[Kovbo]

Summary

• stop importing ART's Unsloth training stack from art.megatron.service at module import time, so Megatron service helpers can load in Megatron-only environments
• keep create_identity_lora usable without TRL/vLLM side imports by using a local GC/CUDA cleanup helper
• remove the Unsloth/unsloth_zoo dependency from SFT tokenization by adding ART-owned response-only label masking in art.preprocessing.response_masking
• centralize chat-template thinking defaults and make the helper types compatible with ty when kwargs are passed to tokenizer.apply_chat_template

Testing

• uv run --with ruff ruff check src/art/megatron/service.py
• uv run ruff check src/art/utils/chat_template.py src/art/preprocessing/tokenize.py src/art/preprocessing/response_masking.py tests/unit/test_preprocessing_tokenize.py
• uv run ty check src/art/utils/chat_template.py src/art/preprocessing/tokenize.py src/art/tinker/server.py
• PYTHONPATH=/tmp/art-serverless-fix/src:/tmp/art-serverless-fix/tests uv run pytest /tmp/art-serverless-fix/tests/unit/test_preprocessing_tokenize.py -q
• PYTHONPATH=/tmp/art-serverless-fix/src uv run python - <<'PY'
from art.megatron.service import create_identity_lora
print(create_identity_lora.__name__)
PY
• created identity LoRA for Qwen/Qwen3-30B-A3B-Instruct-2507 with rank 1