Offload Prefill to CPU/GPU

[ayf7]

Currently, the prefill stage is done through the NPU. Because the NPU requires a static shape and the decode stage is accepts one token at a time, prefill going through sequentially is suboptimal.

A possible solution is to offload prefill to a CPU or GPU model, or alternatively a fixed size on the NPU. This would require extending the functionality of the converter to create another model for the prefill stage, and updating LLMBase and Pipeline.