llama : Async DirectIO model loading on Linux (#18012)

llama : Async DirectIO model loading on Linux (#18012)

* 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