Change error computation to converge in less steps

[jumpinthefire]

By computing the initial error using the residuals of the initial values, and computing the newest error as soon the parameters are updated, we are able to reduce the number of iterations required for convergence.

This PR also:

• Fixes an unintended regression with residual-only computation: par_iter had been turned into iter in my last commit 😓
• Removes some debug logs I'd temporarily introduced and meant to remove.