LLM Inference State of the Art

This is a placeholder issue, will fill out over the next few days with known results (both CPU and GPU) for LLM-inference under FHE, primarily focusing on BERT-base and Llama 7B/8B as most papers seem to evaluate on one of them.

* CHEDDAR: https://github.com/scale-snu/cheddar-fhe/commits/main/
* Theodosian (cheddar follow-up): https://arxiv.org/abs/2512.18345
* Cerium: https://arxiv.org/abs/2512.11269