[Feature]: Add llama.cpp integration to TheRock

[catan2001]

Motivation

Add llama.cpp as an external project to TheRock.

llama.cpp offers a lightweight, high-performance implementation for LLM inference, capable of running efficiently on both CPUs and ROCm-enabled GPUs. Including it in TheRock makes it easier for clients to experiment with and deploy language models, accelerating prototyping and evaluation.

Related to: #1449

Technical Details

llama.cpp is managed using the repo_management.py, llamacpp_repo.py and llamacpp_build.py scripts, in a manner similar to how PyTorch is handled as an external project. To execute it, simply run:

python llamacpp_repo.py
python llamacpp_build.py --rocm-dir <by default /opt/rocm> --gpu-targets <e.g. gfx1030,gfx1100>

For a detailed explanation of its workflow, refer to README.md. The project is currently built using python scripts by cloning upstream repo.

Test Plan

• Built TheRock with llama.cpp included.
• Ran a subset of llama.cpp test binaries to confirm successful compilation on Linux.
• TODO: create smoke-tests and run full test suite.

Test Result

• llama.cpp binaries compiled successfully on Linux.
• Basic functionality verified via sample test runs.
• Full validation and benchmarking still pending.

Submission Checklist

• Read and followed the contributing guidelines: [TheRock Contributing Guide](https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests).
• Add llama.cpp managed via Python scripts.
• Add README.md
• rocWMMA integration. Wait for Add rocWMMA to TheRock #1938.
• Completed extended testing on Linux and Windows.

[catan2001]

I changed PR to Draft since currently llama.cpp is integrated with the submodules instead of it being managed via Python scripts.
I will re-open it for review once the script based management is in place.

[catan2001]

I've updated the test scripts and verified that they work correctly on Linux. As part of this, I created a Python script that automates running the integrated llama.cpp tests, which could be useful for CI/CD workflows. The script currently skips test-backend-ops due to a floating-point exception issue on GFX1030 (tested using GFX1031 via HSA_OVERRIDE_GFX_VERSION), but everything runs smoothly on GFX1201 and GFX1100. It also skips all of the tests that require additional downloading since it would slow down the CI/CD pipeline.

I'm still unsure whether we need additional sanity checks. My initial thought was to verify that the built binaries are linked against HIP/ROCm, but that might be overkill. Perhaps running a minimal binary to confirm functionality would be sufficient. Also, I've switched to using the upstream llama.cpp repo instead of ROCm's fork, which was based on a much older commit. Maybe it would be the best if someone could check what I did already.

[catan2001]

I did some changes related to llama.cpp PR. When someone has a chance, I'd appreciate a review.

Thanks in advance!