Use latest ggml from upstream

[MollySophia]

What's changed:

• Correctly compile and inference with latest ggml.
• Regenerated expected-logits-*.bin using old code base, with GGML_SILU_FP16 turned off. (Upstream ggml uses FP32 for SiLU now, while the old one uses FP16 for SiLU by default, which causes some logits difference in tests.)
• Slightly refactored CMakeLists.txt according to llama.cpp, and added option to enable Metal backend.
• Changed the way of offloading layers to gpu using the new ggml backends and scheduler.
• Use completely unmodified ggml submodule.
• Disabled thread sanitizer. It's broken in llama.cpp too.

TODOs:

• Test basic inference
• Test gpu offloading with Apple Metal
• Test gpu offloading with CUDA (And possibily fix any problems)
• Test gpu offloading with other backends (e.g HIP, OpenBLAS) (May not be possible myself :P)
• Update docs and README.md

[MollySophia]

Update: offloading with CUDA doesn't work yet
I'm working on it