ggml-cuda : add defensive VRAM margin to prevent swapping stalls

[demiquartz]

Overview

This PR adds a defensive VRAM margin in ggml_cuda_device_malloc to prevent system instability and computation stalls caused by VRAM overcommit, specifically in WSL2 with ROCm environments.

• 256MB Margin: Subtracted from reported free VRAM to prevent the OS from initiating aggressive RAM swapping.
• 1GB Grace Interval: Implemented for free_caches to minimize the performance overhead of calling cudaMemGetInfo.

Additional information

In environments like WSL2/ROCm (tested with 7900 XTX), exceeding physical VRAM triggers a "swapping loop" where active inference memory is moved between VRAM and system RAM. This saturates the PCIe bus, causing the GPU computation to stall indefinitely and leading to rapid system RAM exhaustion, even if the process remains active. This PR ensures llama.cpp reports a memory allocation error before this stall occurs.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: YES. AI (Gemini) was used in an assistive capacity to verify C++ implementation details, including operator precedence, type safety for size_t calculations, and cross-compiler compatibility. The core logic (Margin and Grace thresholds) and the diagnosis of the swapping-related stall were conceptualized and validated by the human contributor through practical troubleshooting.

Closes #22583

[am17an]

Create an issue with a reproduction first

[demiquartz]

I have created the issue with detailed reproduction steps and environmental information as requested: #22583
This issue includes a screenshot of the VRAM/System RAM behavior on WSL2 and the specific setup script used to reproduce the hang. Since the root cause and reproduction are now clarified, I would like to request a re-review of this PR.