How disabling prefetching should be implemented?

[Nekotekina]

It's currently impossible to disable prefetching. May need to load huge model using mmap, for one reason or another. Any attempt to use one way or another will start with pointless but very long step of reading all weights at startup (and subsequently paging them out by OS). There seems to be no any way to disable it gracefully.

Basically, huge model can run (slowly) via streaming from SSD if I hard-code disable prefetching.

And it's not the best SSD, it can reach better speeds I'm pretty sure. It's not practical, but that's not the point.

[Nekotekina]

Ok, let's see #28055