feat(core): autograd engine (reverse-mode autodiff)

[Celz-Pch]

Goal

Enable gradient computation through a dynamic computation graph.

Tasks

• Variable / Node wrapping Tensor with gradient tracking
• Record operations into a DAG
• backward() that traverses the graph and accumulates .grad
• Backward implementations for add, mul, matmul, sum

Acceptance criteria

• Gradients match numerical differentiation on toy functions
• Inspired by micrograd / tinygrad design

References

• Andrej Karpathy's micrograd
• PyTorch autograd internals