QNN warning seems out-of-place/unnecessary/different

[woodsp-ibm]

I have just started changing tutorials to use the new circuit library functions to create library circuits rather than the deprecated (Blueprint based) classes. This was a work item in #945.

In changing things I notice this warning with QNN use

No gradient function provided, creating a gradient function. If your Estimator requires transpilation, please provide a pass manager.

which you can see in the published docs as these are the tutorials re-run using the latest code. E.g here https://qiskit-community.github.io/qiskit-machine-learning/tutorials/02_neural_network_classifier_and_regressor.html#Classification-with-an-EstimatorQNN and in a few other places in same tutorial and elsewhere when using QNNs

2 observations/comments:

1. Why do we warn about no gradient function, i,e that it;s using its internally create default. It never warned about that before - and still does not in the present code if you use a V1 primitive.

2. Why the transpiler warning too. In the case of the tutorial its annoying to see these since they are irrelevant in this case. I do not see any other algorithms warning, when using an Estimator or Sampler, that pass manager might need to be set - so why do we do so in the QNNs?

[edoaltamura]

The logic is a bit convoluted there. Hopefully, it will become simpler when #973 merges. At this point, it would make sense to review the exceptions, since V1 primitives drop out of the picture.