CUDA: only init NCCL for setups with multi GPU

[EldarBorge]

Overview

Skips NCCL if the setup is single GPU, i.e. only initializes NCCL if GPU count is more than 1.
This reduces VRAM on single GPU setups that was introduced in b8738.

Fixes #21759

Tested locally and reduces VRAM both idle and under load to pre b8738 levels on my setup.

Requirements

• I have read and agree with the contributing guidelines
• AI usage disclosure: Yes, AI was used to bisect the bad commit and to investigate what in that commit caused higher VRAM usage.

[JohannesGaessler]

#21746 should be a better fix.

[EldarBorge]

You are probably right, your fix is too advanced for me to understand! Let's await @ggerganov , but feel free to close my PR if it's not a good fix. 🙂

[eelgaev]

Is NCCL always initialized now? or is it just for --split-mode tensor?

[EldarBorge]

Is NCCL always initialized now? or is it just for --split-mode tensor?

always after b8738 is my understanding