Upstream the mmap and splits loader from llama.cpp

Please upstream the llama-mmap.cpp and llama-model-loader.cpp from llama.cpp. I'd like to have these features in Whisper.cpp (https://github.com/ggml-org/whisper.cpp/issues/631 , whisper-ggml-large-v3.bin = 3.1G) and another third-party project (Dia.gguf = 6.7G). When using ggml as a submodule, it'd be convenient to not vendor those two .cpp files.