Update HIPBLAS CMake

[daniandtheweb]

The old code for building HIPBLAS with CMake didn't work at all with my self-built ROCm sdk so I've updated it based on the llama.cpp implementation.

Some newer commit also causes linking problems on Debian Unstable with undefined references to cuda code using hipcc.

This new CMake code works with the self-built packages and it should keep the compatibility with the typical ROCm install made with the package manager (Ubuntu 22.04 and Debian Unstable have been tested and both work fine).

Debian Unstable still needs to build using hipcc as there's no ROCm's Clang available (or at least I haven't found any).

[daniandtheweb]

Right now I've tested on Debian with the distribution's HIPBlas and everything works fine.

[daniandtheweb]

Setting the CC and CXX to rocm's clang and clang++ on Ubuntu 22.04, the same as llama.cpp's CI, works fine.

[daniandtheweb]

When I try to replicate the build process from llama.cpp with this:
cmake -B build -S . -DCMAKE_HIP_COMPILER="$(hipconfig -l)/clang" -DGGML_HIPBLAS=ON; the build fails while compiling the tests with the error: unrecognized command-line option '--hip-link'.

However building specifying -DCMAKE_C_COMPILER and -DCMAKE_CXX_COMPILER instead of -DCMAKE_HIP_COMPILER results in a working build.