Enabling building llama.cpp using system libggml

[ckastner]

A llama.cpp build could make use of an already installed libggml using the latter's cmake package. This would be the typical case for Linux distributions.

Note: This requires a ggml build where GGML_BIN_DIR is commented out, see 82d9ca4.

# Build and install ggml to some custom path
$ git clone https://github.com/ggml-org/ggml.git && cd ggml
$ cmake -B build -DCMAKE_INSTALL_PREFIX=/tmp/system-ggml -DCMAKE_INSTALL_RPATH='$ORIGIN'
$ cmake --build build && cmake --install build

# Build and install llama against this ggml, using ggml's cmake package
$ cmake -B build \
	-DLLAMA_USE_SYSTEM_GGML=ON \
	-DCMAKE_PREFIX_PATH=/tmp/system-ggml/lib/cmake/ggml \
	-DLLAMA_BUILD_EXAMPLES=ON \
	-DLLAMA_BUILD_TESTS=ON
$ cmake --build build
$ build/bin/llama-bench
...
$ bash ./ci/run.sh ./tmp/results ./tmp/mnt
...

# Negative check, to make sure we didn't break the regular build
$ cmake -B build \
	-DLLAMA_USE_SYSTEM_GGML=OFF \
	-DLLAMA_BUILD_EXAMPLES=ON \
	-DLLAMA_BUILD_TESTS=ON
$ cmake --build build
...
$ bash ./ci/run.sh ./tmp/results ./tmp/mnt
...

The tests passed but I get a ModuleNotFoundError (torch) from convert_hf_to_gguf.py, which is unexpected. I haven't investigated yet.