CUDA: fix typo in FlashAttention code (#13926) #2

Job	Run time
macOS-x64	3m 29s
macOS-arm64	1m 53s
ubuntu-22-vulkan	5m 36s
ubuntu-22-cpu (arm64, ubuntu-22.04-arm)	2m 58s
ubuntu-22-cpu (x64, ubuntu-22.04)	2m 44s
ios-xcode-build	12m 4s
windows-sycl	10m 14s
windows-cpu (arm64)	2m 47s
windows-cpu (x64)	3m 24s
windows-hip (radeon, gfx1100;gfx1101;gfx1102;gfx1030;gfx1031;gfx1032)	49m 27s
windows (opencl-adreno, arm64, -G "Ninja Multi-Config" -D CMAKE_TOOLCHAIN_FILE=cmake/arm64-window...	1m 35s
windows (vulkan, x64, -DGGML_VULKAN=ON, ggml-vulkan)	3m 32s
windows-cuda (11.7)	31m 0s
windows-cuda (12.4)	40m 40s
release	2m 14s
	2h 53m 37s

Provide feedback