`feat`: implement mini-batch training support

[FiberedSkies]

Description

Currently, the training loop in SheafNN::train() processes cochain data one by one, which is both inefficient and unstable as heck. We need to implement mini-batch training support to address this.

Details

• Add batch_size parameter to train() and train_debug() methods
• Create DataLoader utility struct for batching supervised data pairs
• Modify forward pass to handle batched inputs
• Update gradient computation to accumulate gradients across batch
• Add batch-wise loss computation and averaging
• Update optimizer step to handle batched gradients properly