perf(ds4): compute greedy argmax on GPU (monolithic and layer split)

[Graffioh]

Summary

• compute DeepSeek4 greedy argmax in the GGML GPU graph for local/remote layer-split and all-hot monolithic execution
• copy one token ID instead of the full vocabulary logits during greedy decode
• keep full logits for sampling and penalties, with the hybrid MoE path retaining its CPU fallback
• propagate sampler state and make CPU/CUDA/HIP argmax tie-breaking consistent

Validation

• cached, dynamic, and fused output-path coverage on CPU and supported GPU backends
• sampler propagation and request-reset coverage

Stacked on #503.