ggml-openvino: fix test-thread-safety, gemma, Qwen3.5 stateful and test-backend-ops failures

[ravi9]

Fixes OpenVINO backend failures caused by recent upstream changes:

• the mixed token/embd input path from llama: support both embd + raw tokens in batch ggml-org/llama.cpp#29622 (0bb496dbd)
• new test-backend-ops cases from metal : few-row MMA mat-mul and batched copies for speculative decoding ggml-org/llama.cpp#29869 (a3a1c4747)
• new test-backend-ops cases from webgpu: add f16 support to fill/set_rows ggml-org/llama.cpp#29897 (11fe02151)
• the single recurrent state gather in build_rs from graph: fix CI realloc abort by gathering the recurrent states once ggml-org/llama.cpp#29856 (436f6f89e)

Changes

• Skip unselected graph branches and support DUP. llama: support both embd + raw tokens in batch ggml-org/llama.cpp#29622 adds a mixed token/embd branch to every input embedding graph through ggml_build_forward_select(). Its nodes are not flagged for compute, but the backend translated them, and the DUP in that branch was unsupported. The scheduler split the graph and passed the embeddings across the split with a fixed token count, so the first single-token decode failed:
  Can't set the input tensor with index: 15, because the model input (shape=[1,1,6,288]) and the tensor (shape=(1.1.1.288)) are incompatible.
  The OV model is now built from the compute nodes only, and a same-type contiguous DUP is translated like CONT. Both parts are needed: either one alone still fails.
• Make inp_scale_rows token dim dynamic. llama: support both embd + raw tokens in batch ggml-org/llama.cpp#29622 also moves the per-token embedding scale (gemma3, gemma3n, gemma4) into a new [1, n_tokens] input. It had a fixed shape, which broke gemma-3 on every device. It now has a dynamic token dim and is padded per chunk on the static (NPU) path.
• Skip GPU MUL_MAT op tests with unbound Q4_1/Q4_K weights. Op tests build Q4_1/Q4_K weights as u4 with an f16 zero point. The GPU plugin fails to compile that form for some row counts ([GPU] clFinish, error code: -5 CL_OUT_OF_RESOURCES). That aborts test-backend-ops on the MUL_MAT cases added in metal : few-row MMA mat-mul and batched copies for speculative decoding ggml-org/llama.cpp#29869. Model weights use a u4 zero point and are not affected. These cases are reported as unsupported on GPU until the plugin is fixed. The check matches only unbound weights, so model loading keeps Q4_1/Q4_K weights on the GPU.
• Handle the single recurrent state gather of build_rs. graph: fix CI realloc abort by gathering the recurrent states once ggml-org/llama.cpp#29856 changed build_rs to gather all recurrent states with one GET_ROWS on the s_copy leaf and take the ubatch and extra states as views of it. The stateful path matched only the previous form, a GET_ROWS per view of s_copy, so Qwen3.5 failed with stateful execution on CPU and GPU (is_axis_valid(axis, r)). For a single-slot cache, the single gather and its views now map to the existing single-slot cases.
• Create FILL in the output type. The new f16 FILL cases from webgpu: add f16 support to fill/set_rows ggml-org/llama.cpp#29897 produced f32 data, which overran the f16 output buffer.
• Reject CONCAT with a quantized type. The new q8_0 CONCAT case from metal : few-row MMA mat-mul and batched copies for speculative decoding ggml-org/llama.cpp#29869 was translated with dequantized inputs, so the quantized output was wrong.

Assisted-by: Claude

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. For debugging, review, fixes and testing.