[Clean Code] 整改非 Linear Attention 模块的代码扫描告警

背景

代码扫描表中分配给作者的非 Linear Attention 告警，需要在不改变训练语义的前提下逐项整改。此前的“代码风格规范修改”Issue #243 对应 1.0 发布检查，现已关闭；本 Issue 只跟踪本批剩余范围。

范围

• hyper_parallel/components/losses/chunked_cross_entropy.py
• hyper_parallel/distributed/context_parallel/attention.py
• hyper_parallel/models/deepseek_v3/adapter/replacements.py
• hyper_parallel/models/qwen3_moe/adapter/chunk_loss.py
• hyper_parallel/models/qwen3_moe/adapter/distributed/context_parallel.py
• hyper_parallel/models/qwen3_moe/adapter/distributed/context_parallel_async.py
• hyper_parallel/models/qwen3_moe/adapter/replacements.py

本批不包括 GDN/KDA 融合算子、其他 Linear Attention 模块、第三方依赖切换或性能优化；不为迎合扫描规则而修改必须保留的行为。

验收

• 对本批扫描项逐处修复，无法安全修改的项说明原因并申请豁免。
• 保持现有接口、CP 通信与训练行为不变。
• 完成相关已有 CPU 测试、diff 检查，并记录仍需的多卡 NPU 验证。

对应 PR：#860。

Review the listed Python files, beginning with the code-scanning findings in chunked_cross_entropy.py and the context-parallel attention and adapter modules. Run the existing CPU tests and perform a diff check before recording any required multi-card NPU validation. Done means each in-scope warning is safely fixed or its exemption reason is documented, with interfaces, CP communication, and training behavior unchanged.