Docs: consolidate build-from-source and GPU backend guide

该 issue 指向整合构建文档，可能在 docs/build.md 或新的 docs/gpu.md 中。首先，请查看现有的 README、CONTRIBUTING.md 以及链接的 GitHub issues（#2013、#695、#721、#756），以了解当前分散的信息。然后，为每个 GPU 后端创建一个清晰的 CMAKE_ARGS 表格，并添加常见错误的故障排除步骤。通过使用 GPU 后端构建项目来验证更改，确保说明有效。

Problem

The build-from-source documentation is spread across README, CONTRIBUTING.md, and many GitHub issues (e.g. #2013, #695, #721, #756). Users with newer GPUs (RTX 50-series, sm_120) or uncommon CUDA versions repeatedly hit the same build failures because the docs lack:

1. A CMAKE_ARGS mapping table showing which flag to set for each GPU backend
2. CUDA version compatibility notes (e.g., CUDA 12.9+ requires specific cmake workarounds)
3. The minimum/maximum supported CUDA toolkit versions
4. How to verify the build actually used the GPU (simple Python snippet to check backend)

Suggested improvements

In docs/build.md (or a new docs/gpu.md):

| Backend | CMAKE_ARGS                                    | GPU Requirements              |
|---------|-----------------------------------------------|-------------------------------|
| CUDA    | -DGGML_CUDA=ON                                | NVIDIA GPU, CUDA >= 11.3      |
| Metal   | -DGGML_METAL=ON (default on macOS)            | Apple Silicon or AMD GPU      |
| Vulkan  | -DGGML_VULKAN=ON                              | Vulkan 1.2+ capable GPU       |
| SYCL    | -DGGML_SYCL=ON -DGGML_SYCL_TARGET=INTEL       | Intel Arc / Iris Xe           |
| HIP     | -DGGML_HIPBLAS=ON                             | AMD ROCm-compatible GPU       |

And a troubleshooting section for the most common build errors (CMake not finding CUDA, nvcc version mismatch, sm_120 not recognized).

Value

Consolidating this reduces the "can't install with GPU" issues that currently fill the tracker.