BugFix: TorchLayer now supports shot batching

[astralcai]

Context:
TorchLayer raises an error when used with a shot vector.

Description of the Change:

• Adds support for shot batching to TorchLayer
• Fixes bug in _to_qfunc_output_type in qnode.py that turns the entire shot-batched output, as opposed to each output in the batch, into the qfunc output type.

Related GitHub Issues:
Fixes #5319
[sc-58099]

[codecov]

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 99.67%. Comparing base (d03216e) to head (7409dbe).
Report is 4 commits behind head on master.

Additional details and impacted files

@@           Coverage Diff            @@
##           master    #5492    +/-   ##
========================================
  Coverage   99.67%   99.67%            
========================================
  Files         407      409     +2     
  Lines       38019    37867   -152     
========================================
- Hits        37895    37745   -150     
+ Misses        124      122     -2     

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

[dwierichs]

Thanks @astralcai !
Mostly looks good to me.
Before approving, we should figure out what is going on in finite diff and SPSA with the permuted axes.
Also I had a question on a missing usefixtures that does not seem to affect the test passing. Am I confused about how this works? 😅

[mudit2812]

Left a couple of initial comments. Seems to me that this is currently blocked by the expected shape of the returns.

[astralcai]

what is going on in finite diff and SPSA with the permuted axes

@dwierichs Can you elaborate?

[dwierichs]

@astralcai I was just referring to the output shape question for these gradient methods.
I do agree that there seems to be a discrepancy between the output shapes produced by PennyLane and the technical specification doc. But I am a bit surprised that this only surfaces in two specific gradient method tests and it is a bug fix unrelated to the original PR. Might be nice to split this off and e.g. add forward pass tests that detect this difference.

[astralcai]

@astralcai I was just referring to the output shape question for these gradient methods. I do agree that there seems to be a discrepancy between the output shapes produced by PennyLane and the technical specification doc. But I am a bit surprised that this only surfaces in two specific gradient method tests and it is a bug fix unrelated to the original PR. Might be nice to split this off and e.g. add forward pass tests that detect this difference.

This only happens when the measurements are in a list. If you change the list to a tuple, the error does not appear, which is the case for most tests out there. More specifically, this only occurs when you have a list of only one measurement returned from a qnode, which I assume doesn't happen much anywhere else in the tests.

[dwierichs]

Ahh, I understand now, thank you! 🙏 Sorry, I overlooked that distinction.

I still think we should test a forward-pass output shape not only within differentiation tests :) Would you mind adding a tiny test for this? If the code is ever changed, it would be confusing not to have any forward pass tests fail, but two differentation tests.

[dwierichs]

Thanks @astralcai 🎉
Just small suggestions to wrap up, but looks good to me!