[NVIDIA GPU] Optimize collective matmul loops when contracting dim is sharded

[NVIDIA GPU] Optimize collective matmul loops when contracting dim is sharded #5094