sync : llama.cpp (CUDA ReLU, CPU-only with CUDA, bloom fix, etc)

[ggerganov]

Also reverts the CUDA memory pull stuff, which was synced last time, but was later reverted in llama.cpp

copilot:all

[FSSRepo]

@slaren I think ggml_init_cublas() should be removed from ggml_init(...) by now. I was creating a model converter for stable-diffusion.cpp. The only thing I need is the gguf module, but it requires me to create a ggml_context to add tensors to the gguf_context. However, when it compiles with GGML_CUBLAS=ON, ggml_init calls ggml_init_cublas unnecessarily, leading to the error mentioned in this comment.

[ggerganov]

With this change, you can now run with CUDA_VISIBLE_DEVICES=-1 and ggml_init_cublas() should do nothing in this case

[slaren]

Once we move llama.cpp to ggml-backend, we will be able to remove the ggml-cuda calls from ggml.c and make the backend more self-contained, but as it is now it is required to support the "old" way to use ggml-cuda. But as @ggerganov says it, it should be possible to use a CUDA build for CPU-only after ggerganov/llama.cpp#3946