fix(autodiff): accumulate repeated multiplication inputs

[AlekseiChirkovVention]

Summary

• Fix multiplication VJP generation when both operands reference the same input.
• Preserve both partial derivatives and combine them with a symbolic addition.
• Keep the existing VJP result contract and broadcast-reduction behavior unchanged.
• Add structural and numerical regression coverage for mul(x, x).

How it works

Each multiplication operand produces its own partial gradient and completes any required broadcast reduction independently. When both operands share one value ID, the two resulting gradient IDs are combined with an AddOperator, whose output becomes the sole exported gradient for that input.

How to test

1. Run:

   python -m pytest py/tests/test_autodiff_elementwise.py py/tests/test_autodiff_matmul.py py/tests/test_autodiff_reductions.py -q

2. Confirm all 55 focused tests pass.
3. Verify the repeated-input regression generates two multiplication nodes followed by one addition node.
4. Verify numerical execution returns 2 * seed * value.