feat: add CSA + HCA block-pool attention compression (DeepSeek V4)

[pbertsch]

Summary

• New tqai.csa_hca module: pure-torch math primitives implementing the DeepSeek V4 interleaved attention design (mild block pooling for query-dependent top-k selection + aggressive block pooling for a global summary, joined under a single softmax).
• New CSAHCAStrategy registered as "csa_hca" in the pipeline registry; compresses a KV tensor into both views and routes both through the configured quantizer.
• 28 unit tests + a real-model E2E script (scripts/csa_hca_e2e.py) validating against Qwen2.5-0.5B.

Validation

• Sanity (cos = 1.0000): degenerate config csa_m=1, top_k=S, hca_m=S recovers full SDPA exactly.
• Real Qwen2.5-0.5B layer 8, 75-token prompt: at 4× compression cos = 0.96, at 5.8× cos = 0.92.
• Tests: 28 new tests + 786 existing tests all passing.

Test plan

• Hand-calculated reference values for 8-token toy KV
• Block pool / score / top-k / gather primitives covered with edge cases (partial remainder block, block_size > S, GQA)
• End-to-end equivalence with full SDPA in degenerate config
• Strategy compress / decompress roundtrip preserves shape
• Strategy registry lookup
• Real-model Q/K/V capture and cosine comparison (scripts/csa_hca_e2e.py)
• Follow-up: replace mean pooling with learned softmax-gated pooling (matches DeepSeek paper more closely)
• Follow-up: fused MLX/Metal kernel path for the combined attention (tqai.attention_fused)
• Follow-up: substitute csa_hca_attention into a model attention layer and measure generation perplexity

🤖 Generated with Claude Code