ollama - 💡(How to fix) Fix Fuse Qwen3Next Gated Delta Net autoregressive decode [1 pull requests]

Qwen3Next and Qwen35 use Gated Delta Net recurrent layers. In autoregressive decode, the current graph expands this recurrent update into many smaller tensor operations. llama.cpp has since added a fused ggml_gated_delta_net op for this path.

This issue proposes bringing that fused autoregressive decode path into Ollama while keeping behaviour unchanged for unsupported backends.

Proposed shape:

• Backport the upstream ggml Gated Delta Net op from llama.cpp while Ollama's pinned llama.cpp commit predates that implementation.
• Use the fused op only for autoregressive decode.
• Keep chunked prefill on the existing explicit graph.
• Gate the fused path by backend capability, so unsupported backends continue to use the existing path rather than silently falling back to CPU.
• Add CPU reference tests, CUDA parity tests for supported head dimensions, and multi-step recurrent state tests.

The change should not add a user-facing option. It is an internal execution optimisation for existing Qwen3Next and Qwen35 models.

Validation expected before merge:

• ggml CPU reference parity.
• CUDA versus CPU parity for supported S_v values.
• Multi-step recurrent state continuity.
• Native ggml, ggml-cpu and ggml-cuda builds.
• End-to-end smoke or benchmark evidence on a Qwen3Next or Qwen35 GGUF.