#730 WARNING: the RAM budget (--resident-budget-gib) cannot be kept - experimental Q4

Description

Hey!

When I try to load the experimental Q4 unsloth model, I see the following strange behaviour: \
VRAM gets filled up (\~15 GB), but RAM doesn't (\~ 0 usage).

Any expert missing in the GPU memory is streamed from the SSD. \
Therefore the performance is deadly slow, about or even less than 1 t/s. \
IQ3 works fine with exactly the same settings, ~50 t/s.

I suppose the main reason is around this warning:
```
FileExpertSource: available RAM 73.21 GiB, 73.31 GiB after the mapped experts left the process working set (#467)
FileExpertSource: RAM budget 66.00 GiB: 22569 of the 65.88 GiB of experts the GPU cache does not hold, by profile rank; the rest are read from the files
FileExpertSource: allocating 65.88 GiB page-locked cache complement
strata generate: WARNING: the RAM budget (--resident-budget-gib) cannot be kept (FileExpertSource: pageable resident complement allocation failed); every expert the GPU does not hold is read from the model files through the OS file cache (--mmap-experts), which is slower
```

I've seen a few similar issues, which are already closed. \
But I'm running the latest version on Oct, 4 and still facing the problem.

The config and log files with absolute paths replaced with `...`:
[strata-unsloth-ud-q4_k_xl.json](https://github.com/user-attachments/files/33020032/strata-unsloth-ud-q4_k_xl.json)
[strata-unsloth-ud-q4_k_xl.log](https://github.com/user-attachments/files/33020031/strata-unsloth-ud-q4_k_xl.log)

System specs: Windows 11, R9 9900X, 96 GB, 6900 XT 16 GB.