tflow BFGS optimizer doesn't update model parameters if it doesn't converge

If you use the BFGS optimizer in tflow, and the calculator doesn't converge (for example, it hits the max steps threshold), the parameters don't get updated in the model.

For example, you train for 10000 steps but BFGS doesn't converge, the model still behaves like it did before the 10k steps.

This was pretty easy to fix, but need to take some time to merge the fix with the most recent updates.