Add batched generation support to model/ package

Fix CPU inference; add KV cache, FIM example, HumanEval results