ggml-cuda: use per-thread stream for buffer-init padding memset

[harkgill-amd]

Overview

ggml_backend_cuda_buffer_init_tensor utilizes cudaMemset for zero tensor padding. This runs on the legacy stream which can collide with parallel HIP graph captures on separate streams causing failures.

This switches it to cudaMemsetAsync + sync (matches existing precedent in the file), keeping it off the legacy stream.

Additional information

Resolves the test-backend-ops -j failures below on ROCm.

ROCm 10.2 + gfx1151 running test-backend-ops -j $(nproc),

• Before

2026-09-11T14:23:16.3067027Z ROCm error: operation failed due to a previous error during capture
2026-09-11T14:23:16.3067555Z   current device: 0, in function ggml_cuda_graph_evaluate_and_capture at /scratch/actions-runner/_work/llama.cpp/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:4366
2026-09-11T14:23:16.3068105Z ROCm error: operation would make the legacy stream depend on a capturing blocking stream
2026-09-11T14:23:16.3068441Z   hipStreamEndCapture(cuda_ctx->stream(), &graph->graph)
2026-09-11T14:23:16.3068935Z   current device: 0, in function ggml_backend_cuda_buffer_init_tensor at /scratch/actions-runner/_work/llama.cpp/llama.cpp/ggml/src/ggml-cuda/ggml-cuda.cu:770

• After

  ...
  Backend 1/2: ROCm0
    ... (all ops OK) ...
  2/2 backends passed
  [exit 0]

Serial case is not affected, passes before and after fix.

Requirements

• I have read and agree with the contributing guidelines: Yes
• AI usage disclosure: Yes, for debug and testing.