fix: DataLoader zero-copy contract violation

[ulises-jeremias]

datasets/dataloader.v docstring claims "Zero-copy: batch extraction uses tensor.slice()
which returns a view", but batch() only uses the view path when batch_indices.len == 1.
When len > 1 it allocates vtl.zeros and copies element-by-element — directly
contradicting the documented design.

Task

Implement true strided batching via slice() + stack() or update the documentation
to accurately describe the current behavior.

Files

• datasets/dataloader.v:33-58