Add ROCm detection and rocm variant support

[leonardopereira10]

#208 Detect AMD/ROCm GPUs and expose a rocm variant. On Linux, detection checks for /opt/rocm or /usr/lib/rocm. On Windows, dxdiag is invoked and parsed for Adapter Family gfx identifiers (supported gfx list added) to identify AMD GPUs; temporary dxdiag_output.txt is cleaned up and protected with try/catch. Also added a Windows rocm asset name in getAssetPattern and added the 'rocm' option to the inference runtime UI.

Description

I included the autodetect implementation, but I should at least keep the option to choose ROCm as the backend, as it offers better stability in contexts with more than 20,000 tokens.

Autodetect is based on a parsing of Windows DxDiag to identify the supported GPU for ROCm (as per the list provided in the HIP documentation).

The download and extraction of HIP versions from llama.cpp has been adjusted by adding the item to the dictionary.

An option has been added to the backend configuration screen to allow selecting llama.cpp HIP/ROCm.

Related Issues

Closes #208

---

Contributor License Agreement

• By submitting this pull request, I confirm that I have read and fully agree to the Contributor License Agreement (CLA), and I am providing my contributions under its terms.

Note

Deleting the CLA section will lead to immediate closure of your PR and it will not be merged in.

[leonardopereira10]

I included the autodetect implementation, but I should at least keep the option to choose ROCm as the backend, as it offers better stability in contexts with more than 20,000 tokens.