fix(gpu): enforce tighter bounds on compression output

[pdroalves]

Compression output should be num_glwe * glwe_mask_size + num_lwes.

closes: https://github.com/zama-ai/tfhe-rs-internal/issues/918

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

[pdroalves]

@tmontaigu I hope this PR fixes the issues you pointed in https://github.com/zama-ai/tfhe-rs-internal/issues/918.

[IceTDrinker]

@pdroalves there was a question on our end about it, do you really need to do the packing on GPU ? we could imagine changing the storage format from CPU to GPU, as this seems to just add complexity for you with no tangible benefits (apart from lower VRAM usage)

[pdroalves]

Performance wise, no. Packing is probably too simple to present any speedup on the GPU (although, I haven't measured that).

Locality wise, yes. We need that to keep compression result in the GPU. I could move data back to the CPU, apply CPU's packing, and move back to the GPU but that would be much more expensive.

[IceTDrinker]

Performance wise, no. Packing is probably too simple to present any speedup on the GPU (although, I haven't measured that).

Locality wise, yes. We need that to keep compression result in the GPU. I could move data back to the CPU, apply CPU's packing, and move back to the GPU but that would be much more expensive.

no but I mean the GPU data would never be serialized, and when you decompress you have to unpack today, isn't that wasteful ?