Refactor tensor arithmetic logic into backend modules

[tran-go]

Problem

CTArray currently mixes user-facing API methods with backend arithmetic logic. Some operations use __tensor_function__, while others call arithmetic operations directly inside the class.

Proposed change

Keep NumPy-style wrappers such as a.sum() inside CTArray, but move the actual arithmetic logic to backend modules such as matrix_arithmetic.py.

Expected structure:

CTArray.sum(...)
    -> __tensor_function__("sum", ...)
    -> registered backend function
    -> OpenFHE implementation