[BUG] TensorFlow version of the backprop tutorial is producing incorrect benchmarks

[josh146]

The time scaling plot in the backprop tutorial is currently incorrect:

Strangely enough, the first section of the demo, comparing single forwards and backwards pass times for parameter-shift vs. backprop, seems to still be returning sensible times.

[josh146]

Update: this looks to be a TensorFlow specific issue, see PennyLaneAI/pennylane#1244. Swapping this demo over to Autograd reverts to the previous timing behaviour.

[josh146]

More info:

The last build in which this demo was correctly passing occurred 8 days ago with build 2968, with demo artifact available here.

For longevity, this build had the following output:

Forward pass (best of 3): 0.005797430699749384 sec per loop
Gradient computation (best of 3): 2.5100458409004203 sec per loop
>>> print(2 * forward_time * params.size)
2.0870750519097783

[josh146]

Update: the demo has been swapped to use autograd, while we investigate this issue.