ggml-cuda: missing __syncthreads() in ssm-scan.cu

[Commit 43e6301 ](https://github.com/ggml-org/ggml/commit/43e6301b01331a7e4c8ee16dc252516dd58cc9e7)
The above commit remove the second __syncthreads() in ssm_scan_f32 on line 81. It seems that this sync is a must or  some threads may read data from smemB/smemC of next loop. 
