ggml-opencl currently stores the GPU buffer in ggml_tensor::data, and after the GGUF changes this will result in a memory leak when not using mmap, as the address of the CPU buffer is lost after the call to ggml_cl_transform_tensor:
https://github.com/ggerganov/llama.cpp/blob/47068e517004d90f13c16352bb3b4cafd53a00cd/llama.cpp#L1516-L1523

This change solves that issue and a possible interaction with ggml-alloc if the opencl buffer falls within the measure buffer memory range by storing the GPU buffer in ggml_tensor::extra instead.

Fixes #2993