19fb69e2ed
Allows opt-in to memory efficient GPU shader at proscribed sequence length. Otherwise, utilizes aggregate MLX primitives for best latency.