But they cause a mysterious bug in test_gpu_fusion on my mac in which the second
of cases 2 and 3 (any order) fails to update the result buffer. I tried but
failed to fix it so far.

Adding tag traces is expensive. This PR disables adding tag traces during the
optimization process itself, which is anyway not the point of tag traces. This
saves 10 seconds of compilation on a graph I'm working with that ends up with
about 240 nodes.

Removed debug print that crept in

Typos

Better error msg.

Fix
Looks fine, mostly it is pep8 compatibility fixes.
Razvan

config flag for max_use_ratio in Equilibrium optimizer

Fix a bug of grad reported by Michael Forbes

The bug is that we do not clip the sequences to n_steps, meaning that when
we revert them for computing the gradient we mess up the gradient.

Pep8
I've looked over it and everything seems fine.