perf: async prefetch of next segment's params during compute

[fszontagh]

Summary

When --stream-layers runs a multi-segment plan, today each merged segment H2Ds its params before compute, then waits. GPU sits idle during the H2D.

This PR overlaps them: while segment N's kernel runs on the runtime backend's stream, segment N+1's params are copied to a new pending buffer (ggml_backend_tensor_copy on cudaStreamPerThread). At the next iteration offload_partial_params recognizes the prefetched signature and adopts the pending buffer in place of a second H2D.

Per-segment wallclock drops toward max(H2D, compute) instead of H2D + compute. Falls back to sync if the pending allocation fails.

Related

Continuation of #1576, #1598, #1601, #1611, #1612.

Numbers

RTX 3060 12 GB, --offload-to-cpu --stream-layers --max-vram -1:

Workload	Before	After
SDXL bf16 1152x896 batch=2 8 steps	20.9 s	18.7 s
Z-Image bf16 1024x688 batch=2 9 steps	82 s	54.7 s

SDXL is a 1-segment plan so prefetch is a no-op; small win is from compute_async + synchronize having less host overhead than synchronous compute. Z-Image hits the 9-segment streaming path and gets the full overlap.

Checklist

• I have read and confirmed this PR follows the contribution guidelines.