llama: name fused GDN outputs before callbacks

[bogdanr]

Summary

Name fused GDN result tensors before running the graph callback.

Context

The FGDN support check looks for graph nodes named with the __fgdn_ar__-<layer> / __fgdn_ch__-<layer> prefix. With Gemma E2B I saw that check hit a generic name (node_47) on the CPU path, which made the support check fail before the fused path could be used.

The callback already gets the same name, but this makes the name explicit on the ggml_gated_delta_net result before the callback runs.

Notes

This did not show a measurable CPU speedup in my local Gemma E2B benchmark, and Vulkan was already enabling the fused path in my setup. The reason for the change is to keep the tensor naming consistent for the support check, not to claim a performance win.

Testing

cmake -S . -B /tmp/llama-pr-build \
  -DLLAMA_BUILD_TESTS=OFF \
  -DLLAMA_BUILD_EXAMPLES=OFF \
  -DLLAMA_BUILD_SERVER=OFF \
  -DGGML_VULKAN=OFF \
  -DGGML_CUDA=OFF \
  -DGGML_METAL=OFF
cmake --build /tmp/llama-pr-build --target llama -- -j2

I also checked the patch in a local llama-cpp-rs build with Gemma E2B. Fused GDN enabled cleanly there, but I did not observe a meaningful performance change.

[bogdanr]

Hi @CISC, small nudge on this one. I've just rebased it onto latest master so it's clean and ready to go. It fixes a crash when loading hybrid Gated Delta Net models (Gemma E2B) on the CPU path, where the fused GDN check hits an unnamed node and aborts. Would really appreciate a review whenever you get a minute. Thanks!

[CISC]

#24646