* Uncached model read
* Removing additional --mmap arg
* Removing trailing whitespaces
* Adding fallback when O_DIRECT is not supported
* Remove branching in llama-model-loader.cpp and reduce code duplications in llama-mmap.cpp
* Adding maybe unused keyword for Mac and Windows.
* File seek aligned
* Removing all branches for direct_io in llama-model-loader.cpp
* Always use alignment from llama_file
* use_mmap=true
