Add DSA (DeepSeek Sparse Attention) principle documentation - #1

GLM-5 integrates DeepSeek Sparse Attention (DSA) to cut long-context deployment cost, but the repo had no explanation of how it works. This adds a self-contained reference document.

Changes

• New doc docs/DSA原理详解.md — Chinese-language explanation of DSA covering:
  • Motivation and the O(L²) bottleneck of dense attention at 1M-token context
  • Two-stage design: Lightning Indexer (cheap, FP8-friendly relevance scoring) → top-k token selection → precise attention only on selected tokens
  • Interplay with MLA (KV-cache compression vs. per-step compute reduction)
  • Complexity/memory comparison table (O(L²) → ≈ O(L·k))
  • Dense→sparse continued-training recipe
  • GLM-5 usage and GLM-5.2 IndexShare improvement (shared indexer per 4 layers, ~2.9× per-token FLOPs reduction at 1M)
  • Trade-offs, use cases, and references (technical report, IndexShare, Transformers glm_moe_dsa)

Notes

• Documentation-only change; no code paths affected.
• Some quantitative/implementation specifics are attributed to official reports, with a closing note directing readers to authoritative sources.