chore(gpu): rework single carry prop to avoid using local streams

[agnesLeroy]

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]

Multi-GPU tests running here: https://github.com/zama-ai/tfhe-rs/actions/runs/12393147259/job/34594035478

[agnesLeroy]

On 1xH100 there does not seem to be any effect on the latency & throughput benchmark measurements.

• ERC20 bench with this change: https://zama.grafana.net/d/ee09fw09s4h6oa/erc20?orgId=1&from=now-30d&to=now&timezone=browser&var-hardware=n3-H100x1&var-branch=al%2Frework_single_carry_prop
• On main: https://zama.grafana.net/d/ee09fw09s4h6oa/erc20?orgId=1&from=now-30d&to=now&timezone=browser&var-hardware=n3-H100x1&var-branch=al%2Ffix_lut_size

[guillermo-oyarzun]

This is change is a bit complex. Just to be extra safe, it should be verified with: add, unsigned_overflowing_add, signed_overflowing_add, unsigned_overflowing_scalar_add, just to test all combinations

[agnesLeroy]

All multi-GPU tests have passed, I also ran those functions locally they were passing.

[guillermo-oyarzun]

All multi-GPU tests have passed, I also ran those functions locally they were passing.

ok, then good to go!