Add: SPMD paged attention with TPUSH/TPOP MIX kernel pipeline

[chenshengxin2026]

Summary

• Add scene test for SPMD paged attention using TPUSH/TPOP inter-core communication between AIC (Cube) and AIV (Vector) cores
• Implements cooperative FlashAttention-style pipeline: AIC handles QK/PV matmuls, AIV handles online softmax and output accumulation
• Uses QK-first / SF-first software pipelining to overlap Cube matmuls with Vector softmax across KV block iterations
• Hardware block_num fixed at 24 with stride-loop distribution over batch * q_loop logical blocks
• Supports q_tile=16 and q_tile=64 static dispatch, adapting to num_heads at runtime

[gemini-code-assist]

Code Review

This pull request implements a Paged Attention MIX kernel using AIC and AIV cores with TPUSH/TPOP communication, including orchestration logic and integration tests. The review identifies critical risks of 32-bit integer overflow during pointer arithmetic in the QK and PV steps when handling large KV caches. Additionally, it is recommended to replace the inefficient scalar loop used for zeroing the destination buffer with a vector-based approach to improve performance on the AIV core.