CUDA: fuse rope + set_rows

[am17an]

Based on #16769.

On a 4090:

Model	Test	t/s master	t/s cuda-rope-fusion	Speedup
llama 8B Q4_K_M	tg32	134.90	136.07	1.01
llama 8B Q4_K_M	tg64	131.41	132.84	1.01
llama 8B Q4_K_M	tg128	130.54	131.87	1.01
qwen3moe 30B.A3B Q4_0	tg32	167.18	168.23	1.01
qwen3moe 30B.A3B Q4_0	tg64	161.00	161.90	1.01
qwen3moe 30B.A3B Q4_0	tg128	158.84	159.83	1.01

[ORippler]

While the fusion itself is quite simple, I would still recommend to add a test to test-backend-ops for it nonetheless

[am17an]

While the fusion itself is quite simple, I would still recommend to add a test to test-backend-ops for it nonetheless

There is already a test added in the vulkan PR #16769

[JohannesGaessler]

I would suggest you use ggml_cuda_cast defined in convert.cuh. Otherwise there will potentially be issues with FP16 <-> BF16 conversions.

[JohannesGaessler]

When you're already working on optimizing RoPE: I think the memory access pattern here is suboptimal because there are gaps between each thread and I don't know whether the compiler is smart enough to combine the first and second write into a single one. I would suggest grouping the values as float2/half2 and either casting dst to that type or using ggml_cuda_memcpy_1.