Add info about CUDA_VISIBLE_DEVICES

[SlyEcho]

Add a sentence about GPU selection on CUDA.

Relevant: #1546

[roperscrossroads]

@SlyEcho

Does this actually work? I struggled with this a few days ago using a slightly older version of llama.cpp. It kept loading into my internal mobile 1050ti and running out of memory instead of using my 3090 (egpu). I was doing something like this:

CUDA_VISIBLE_DEVICES=1 ./main -ngl 60 -m models/model.bin

It always went to the internal GPU with id 0.

Tomorrow morning I will give it another try, but I do not think it works if you set the ID from nvidia-smi (in my case: 1050ti: 0, 3090: 1).

[JohannesGaessler]

CUDA_VISIBLE_DEVICES does work on my test machine using the master branch. In any case, it should be possible to just control this via a CLI argument before long. My current plan is to add something like a --tensor-split argument for the compute-heavy matrix multiplication tensors and a --main-gpu argument for all other tensors where multi GPU wouldn't be worthwhile.