l2 regularization of weights only

[Engineero]

Modified l2_loss term so that it does not count biases, only weights. Typically we do not want to regularize biases as they are:

1. additive, not multiplicative, and so are far less likely to cause numeric issues if they get too large, and
2. not necessarily limited by real-world constraints to be near zero.

Basically our data can have biases that are far from zero, so we shouldn't try to keep them close to zero.