feat(opencl): add optional OpenCL backend support via Cargo feature

[admiralakber]

Hey mate,

Simple yet high value add, I thought I should upstream it. I have tested this on Linux and Android. The OpenCL backend is much more reliable than Vulkan.

✨ Summary

• Introduce an opencl feature for both crates, package ggml-opencl sources, and enable GGML_OPENCL in build.rs with per-OS linking to the OpenCL loader. Keeps changes consistent with existing CUDA/Vulkan patterns.

🔎 Diagnosis

• Place behind rust feature flag to avoid breaking existing builds
• Grab defaults from llama.cpp upstream: https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/OPENCL.md

🚀 Changes

• New feature flag on llama-cpp-2/Cargo.toml and added new globbing to llama-cpp-sys-2/Cargo.toml
• Little tweak to the llama-cpp-sys-2 README.md should help SEO if that matters anymore
• Most of the work in llama-cpp-2-sys/build.rs - matched the existing patterns

☁️ What you’ll need to do

• Just build with the opencl feature flag

🧪 Smoke test

cargo run -p llama-cpp-2 --example usage --features opencl -- ${PATH_TO_MODEL} --verbose 

✅ Acceptance

• If someone else can do a manual test that would be great. A Linux box with a dediated GPU would be strong validation.
  • Note: It might be that the llama.cpp OpenCL backend is only supported for Android which I can report is working great.
• A Windows or Mac build just to confirm the rust link commands in build.rs are correct.