cuda: add GGML_CUDA_BLOCKING_SYNC to eliminate 100% CPU busy-wait on full GPU offload

[hksk]

Overview

When running llama-server with full GPU offload (--n-gpu-layers 99), the host thread that drives decoding ends up busy-waiting inside cudaStreamSynchronize. This makes one CPU core sit at 100% even though the GPU is doing all the work.

This PR adds an opt-in environment variable GGML_CUDA_BLOCKING_SYNC. When set to a non-zero value, llama.cpp calls cudaSetDeviceFlags(cudaDeviceScheduleBlockingSync) before the CUDA primary context is created, turning the busy-poll into a blocking sleep and dropping host CPU usage to near zero during decode. Throughput stays essentially the same; the only expected cost is a small increase in host-GPU synchronization latency.

Additional information

• Related discussion that validated this fix: 100% CPU single core usage while everything is fully offloaded to the GPU #22238.
• Tested on a Proxmox LXC running Debian 12, with an NVIDIA Quadro P5000 (Pascal, sm_61), CUDA 12.6, using the Qwen3.5-9B-UD-IQ2_M.gguf model.
• Both a heavy request (1000 tokens) and a long-context request (~31k tokens) completed successfully without CPU saturation.

Requirements

• I have read and agree with the contributing guidelines.
• AI usage disclosure: YES — I used an AI assistant to help draft the PR description and the initial patch. I reviewed, tested, and verified the changes on my own hardware.

[ggml-gh-bot]

Hi @hksk, thanks for your contribution!

Per our contribution guidelines, the automated PR checker found the following issue(s) that need your attention:

• PR Template not respected: Please respect the template when creating a new pull request. Make sure to fill out all required sections.

• AI-generated content: While code is allowed to be generated by AI, please write the PR description and commit messages on your own without the help of AI.

---

Please note that maintainers reserve the right to make final decisions on PRs. If you believe there is a mistake, please comment below.