feat: sync llama.cpp

[jhen0409]

• ggml-alloc : correctly check mmap return value for errors ggml-org/llama.cpp#3075
• Posixify madvise and pagesize. ggml-org/llama.cpp#3037 breaks the Android build, it can be fixed by bump NDK version to v25. Choice to temporary use a patch for keep working for some Android project that still using NDK 24. I didn't to open a PR to llama.cpp because NDK 24 is obsoleted we should deprecate it.