refactor(gpu): fix sample extraction when nth > 0 and keep input unchanged

[pdroalves]

This PR:

• introduces a rust test for GPU's sample extraction,
• fixes it when nth > 0,
• refactors it so the input (the glwe) is not modified by the method.

We might also observe another performance improvement.

closes: https://github.com/zama-ai/tfhe-rs-internal/issues/671

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]

Hey @pdroalves! Thank you very much for this. I have some comments on the PR but mainly a question: shouldn't we apply sample extraction to a list of GLWE ciphertexts and output a list of LWE ciphertexts? That would be more efficient for the compression of a vector of inputs, which is what we'll need if I'm correct. Then you don't need to add types for CudaGlweCiphertext and CudaLweCiphertext, you can reuse the existing List types.

[agnesLeroy]

Hey @pdroalves! Thanks a lot for this PR 🙂 I just have some small comments about it, after that it'll be good to go!

[agnesLeroy]

All good! Thanks!