- modify aligned_matmul kernel for dynamically malloc memory
- fix top_k_avg_pooling kernel to support data whose size is more than cuda shared memory.