fix: filter thinking tags from non-SSE local provider streams

[ming1523]

Summary

• Filter reasoning blocks from non-SSE JSON streams from local LLM providers.
• Share the incremental thinking-block parser between SSE and non-SSE streaming.
• Add regression coverage for complete, split, and unclosed thinking tags.

Tests

• python3 -m pytest -q tests/services/llm/test_local_provider.py
• python3 -m compileall -q deeptutor/services/llm tests/services/llm/test_local_provider.py
• python3 -m ruff check deeptutor/services/llm/local_provider.py tests/services/llm/test_local_provider.py