Hybrid model loading with DirectIO and MMAP

[JTischbein]

This is a draft of the hybrid model loading. It combines mmap for host buffers and DirectIO for device buffers.

• Allows weight streaming from disk with the CPU backend
• In case the backend is not capable of asynchronous updating the loading falls back to mmap (allows backends to decide whether it supports DirectIO or not)

Currently only the non-Windows path is implemented. Windows path will follow.