Investigate GPU offload in scheduler

[antoniupop]

No description provided.

[antoniupop]

The use of Multi Bit PBS parameters alongside compression is not well supported due to an inconsistency in the HL API.

Issues linked to this:

• https://github.com/zama-ai/tfhe-rs-internal/issues/898
• https://github.com/zama-ai/concrete-spec/issues/154

Fixed and tested.

[antoniupop]

It is currently not possible to target specific GPUs for execution which restricts the scheduling options and forces multi-gpu usage to parallelize operations rather than across operations.
This has been requested as a feature from tfhe-rs and is now a Q1 ORK: https://github.com/zama-ai/tfhe-rs-internal/issues/889

Fixed/available

[antoniupop]

Performance inconsistencies in ERC20 benchmarks in particular between GPU types (on H100 - with NVLink or SXM).
Ongoing work: https://github.com/zama-ai/tfhe-rs-internal/issues/851

[antoniupop]

Decompression on GPU issue after de-serialization. Details
Investigation ongoing: https://github.com/zama-ai/tfhe-rs-internal/issues/905