feat(gpu): Implement unsigned_overflowing_scalar_add

[bbarbakadze]

PR content/description

Implement overflowing scalar add

Check-list:

• [x] 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

[tmontaigu]

I don't think the first commit is right with regards to how the GPU works compared to CPU

On the CPU, when the input to a PBS is a trivial ciphertext, we actually skip the PBS and apply the LUT directly via clear operations, so the output is still a trivial ciphertext without noise https://github.com/zama-ai/tfhe-rs/blob/e825277219cc142bf13959b486192ddc97711445/tfhe/src/shortint/server_key/mod.rs#L750

I don't think you do that on GPU, meaning that you apply PBS on trivial inputs meaning the output has noise and is no longer a trivial

This is something you should probably also do if possible as it allows using trivial ciphertext to debug fhe operations much faster because the computations are faster (no pbs) and since there is no noise you can use a debuger/print-debug to see values and then figure out what is wrong

[agnesLeroy]

Ok so we'll remove the first commit and in case the gpu feature is activated we won't check if NoiseLevel is zero in case a trivial ciphertext is given as input to unsigned_overflowing_scalar_add

[agnesLeroy]

We just pushed the change

[tmontaigu]

No you should keep that check I think,

Because on GPU you should also be able to see that the clear value being added trivialy overflows

https://github.com/zama-ai/tfhe-rs/blob/e825277219cc142bf13959b486192ddc97711445/tfhe/src/integer/server_key/radix_parallel/scalar_add.rs#L26

[tmontaigu]

Hum ok I looked at the wrong think, I check again