clarify that qnodes with qml.sample are not differentiable

[timmysilv]

Context:
Autograd allows the differentiation of integers, although this doesn't entirely make sense (see #2556). On top of this, our docstring for qml.sample currently suggests that it is possible, when it should really just point the user elsewhere.

Description of the Change:

• Update the docstring of qml.sample to suggest the use of expval instead
• Add tests to show that autograd allows computing the jacobian of a circuit which returns samples, whereas jax (for example) does not

Benefits:
More useful docs

Possible Drawbacks:
We are testing something that shouldn't exist just because Autograd said it should exist

Related GitHub Issues:
Fixes #4840

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.67%. Comparing base (3407a9e) to head (90c0a43).
Report is 1 commits behind head on master.

Additional details and impacted files

@@            Coverage Diff             @@
##           master    #5237      +/-   ##
==========================================
- Coverage   99.68%   99.67%   -0.01%     
==========================================
  Files         399      399              
  Lines       36853    36567     -286     
==========================================
- Hits        36736    36449     -287     
- Misses        117      118       +1     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[trbromley]

Thanks @timmysilv!

[astralcai]

Although it is possible with autograd, should we at least raise a warning saying that what the user is trying to do doesn't necessarily make sense?

[timmysilv]

the trouble with all this is that auto-diff can be rather opaque. we don't generally have the QuantumScripts available to us ahead of time, so we can't inspect the measurements in order to know when to raise the warning. an alternative way to think about this is, QNodes won't know that they're being executed within a qml.grad (or other auto-diff) context. We have to be cognizant of this at all times when modifying the execution pipeline, but sometimes things just don't work, like this bug.