Simplify backward pass by returning local gradients per op

[Lyt060814]

Summary

As suggested by @karpathy in this tweet (a follow-up to the microgpt announcement):

You just return local gradients for each op and get backward() to do the multiply (chaining) with global gradient from loss. So each op just expresses the bare fundamentals of what it needs to: the forward computation and the backward gradients for it.

This PR implements that simplification:

• Each op now stores its local gradients in a _local_grads tuple instead of defining a _backward closure
• backward() uniformly applies the chain rule by multiplying local gradients with the upstream gradient
• _children kept as tuple instead of set() to preserve ordering correspondence with _local_grads

Test plan

• pytest test/test_engine.py — forward and backward results match PyTorch (both test_sanity_check and test_more_ops)
• demo.ipynb — MLP training on make_moons converges to 100% accuracy, loss curve unchanged
• trace_graph.ipynb — computation graph visualization works correctly