Early stopping of square root BP

[mtfishman]

"Square root BP" as defined in Eq. (25) of ^1 is introduced in #100, but without an early stopping criteria. Early stopping is a bit more subtle since in general there is an arbitrary unitary gauge degree of freedom of the "square root message tensors", which could be fixed in various ways, such as by squaring them, using a unique factorization like a positive QR or polar decomposition, using the singular values of the square root message tensors, etc. This is a reminder to implement that.