cuda: fix rope fusion for gemma3

[am17an]

Fix #17322, apparently we weren't checking the order of the ops here and gemma3 has a peculiar set of ops which seem to match ggml_cuda_should_fuse_rope_set_rows

[CISC]

BTW, I think there are several issues in ggml_cuda_can_fuse.

There is a pattern of checking the ops size, but not the content before calling ggml_can_fuse_subgraph with ops, like here (which makes the second call redundant):

llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu

Lines 3084 to 3085 in 01590ff

	if (ops.size() == 5 && (ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}) ||
	ggml_can_fuse_subgraph(cgraph, node_idx, ops, {node_idx + 4}))) {

It should have checked that mul_mat_bias_glu_ops or mul_mat_id_bias_glu_ops were the ops we were trying to fuse first, then made only one ggml_can_fuse_subgraph call.

[am17an]

@CISC yeah, I'm refactoring that function. Will open a separate PR

[bssrdf]

@am17an, I am interested in implementing some fusion ops for sd.cpp ,which should really benefit from it. Is there any guide about how to add these fusion ops in cuda backends? Or just follow the existing examples? I think a general/brief introduction to this topic is helpful for other developers.

[am17an]

@bssrdf Sure, let me write up something