fix(data): supervise historical Qwen thinking turns

[cheater6666]

Motivation

Qwen3.5 and Qwen3-Next-Thinking render historical assistant turns without the <think> scaffold used by the final assistant turn. The parser only matched the longer configured header, silently masking all historical assistant outputs from training.

Modifications

• Allow chat templates to declare alternate assistant headers.
• Match headers longest-first so the final-turn thinking scaffold remains excluded.
• Register historical headers for Qwen3.5 and Qwen3-Next-Thinking.
• Add a focused parser regression test and refresh the real-tokenizer references.

Related Issues

Fixes #912

Accuracy Test

Input IDs are unchanged. All changed loss-mask values are 0 -> 1 in historical assistant spans.

• Qwen3.5 supervised tokens: 146 -> 257
• Qwen3-Next-Thinking supervised tokens: 125 -> 219

python -m unittest \
  tests.test_data.test_assistant_header_alternates \
  tests.test_data.test_parsers.TestTemplatePreprocessing.test_qwen35_instruct \
  tests.test_data.test_parsers.TestTemplatePreprocessing.test_qwen3_next_thinking -v

Result: 3 tests passed.

Benchmark & Profiling

Not applicable; this changes preprocessing correctness, not runtime inference performance.

Checklist

• Ran the repository pre-commit hooks on all changed files.
• Added a unit test and updated real-tokenizer regression references.
• Documentation changes are not needed for this bug fix.
• Performance and model accuracy benchmarks are not applicable.