chore(gpu): allocate pinned host memory when possible

[agnesLeroy]

This speeds up the copy to the GPU

closes: please link all relevant issues

PR content/description

Check-list:

• [ ] Tests for the changes have been added (for bug fixes / features)
• [ ] Docs have been added / updated (for bug fixes / features)
• [ ] Relevant issues are marked as resolved/closed, related issues are linked in the description
• [ ] Check for breaking changes (including serialization changes) and add them to commit message following the conventional commit specification

[agnesLeroy]

Do we know pinned memory will bring any advantage? AFAIK it is more expensive than malloc and only really useful if you repeatedly run copies using the pinned pointer, which I don't think happens often in our case.

My intention was to run the benchmarks to see the effect. I've rebased on main and triggered the multi-bit 64 bit ones we should get the results soon :slightly_smiling_face:

[agnesLeroy]

@pdroalves so this doesn't seem to have any effect on performance: grafana link

[agnesLeroy]

I' running complementary benchmarks. Actually I think this will have some effect when using multiple GPUs, based on this: https://medium.com/gpgpu/multi-gpu-programming-6768eeb42e2c

[agnesLeroy]

@pdroalves so I've checked single & multi-GPU performance on this branch compared to main and there is no improvement. If anything we seem to loose 1ms in the multi-gpu benchmarks for all precisions & operations. I guess we don't want to merge this. single-gpu run: https://github.com/zama-ai/tfhe-rs/actions/runs/9693223295/job/26748315432 multi-gpu run: https://github.com/zama-ai/tfhe-rs/actions/runs/9695743021 (Benchmarks on main are still ongoing, they haven't been uploaded to Grafana with the fix in the FFT yet).

[pdroalves]

Yeah, don't think our case benefits from pinned memory.