Multi-Batch Proof Generation

[ppotash]

Current code assume inference input is a single batch. Add functionality for generating proofs over multi-batch inference.

[ppotash]

If we switch to multi-batch, given how the hooking works, we'll seemingly have to generate the onnx files after each inference pass for a given batch (maybe we multiprocess it to make it more efficient: whenever a lora module has completed its forward pass, the onnx file gets created for it will the rest of the network is still propagating).

[ppotash]

possible there would be two options for inference: 1) single batch for faster inference (onnx files aren't generated until after inference, so there is no slowdown there); 2) multi-batch inference -- slower inference since we have to wait for the the onnx files to be saved after each forward pass, or else the activations will be lost when the next pass occurs.