Add option to control figure generation in composite experiments

[coruscating]

Summary

Closes #1239. An init parameter has been added to composite analysis, generate_figures, to control figure generation. By default, generate_figures is always, meaning figures will always be generated. If generate_figures is set to selective, then only figures for analysis results of bad quality will be generated. If generate_figures is set to never, then figures will never be generated. This behavior can still be overridden for individual analyses by setting the analysis option plot, which is now unset by default instead of True.

Details and comments

• I didn't end up implementing the proposal discussed in the issue where all figures would still be generated in selective mode if there are 10 or fewer figures. If this is preferable, we can add the behavior, but it seems a bit messy to keep a counter of all component experiments and how many figures each generates.
• Added how to access child analysis classes in the Getting Started tutorial.

Timing benchmark

I ran a parallel T1 experiment over 127 qubits then timed only the analysis.

• always: ~120 seconds
• selective: ~100 seconds (generated 60/127 figures)
• never: ~15 seconds

[coruscating]

This chisq > 0 condition has been added to quality=good because of edge cases where zero chisq is a sign that something is wrong (in this case, running RB with only one sample per length)

[itoko]

Thank you for the explanation. This change itself makes sense to me. (I'm not sure we can really determine if a fit is good or bad based on a single value, though, but I know it's another issue.)

[itoko]

Thank you, @coruscating. I think generate_figures is a useful option for CompositeAnalysis.
I have two high level questions.

1. Should we expose generate_figures option in Experiment classes?
2. Do we have any means to generate some figures (without rerunning analyses) after running analyses with generate_figures=="never"?

(1) I think an Experiment class should not depend on an Analysis class as possible (an Analysis class may depend on Experiment classes). Basically, I don't want the changes in Analysis class affect Experiment classes. In this case, asking users to set generate_figures via analysis option makes sense to me. What do you think?

(2) I guess no but I know adding such a feature is beyond the scope of this PR. I hope the on-going refactoring on AnalysisResult makes the future addition of the feature much easier.

[coruscating]

@itoko Thanks for your review!

(1) I was thinking it would be convenient for users to specify generate_figures in the same interface as flatten_results, which is also an option set upon composite experiment instantiation that only affects the analysis and experiment data classes. I agree it would make more sense for both of these options to be analysis options, but changing the behavior of flatten_results would be a breaking API change so I'd rather keep it as is until we do the refactor. What do you think? I can also move generate_figures to the analysis class and keep flatten_results where it is now.

(2) Good question. There's no easy way to do that now since you have to give the plotter the series data manually. I'll write an issue to track this feature.

[itoko]

@coruscating thank you for the quick responses.

Regarding (1), your suggestion (move generate_figures to the analysis class and keep flatten_results where it is now to avoid a breaking API change within this PR) makes sense to me.

[coruscating]

@itoko I've updated generate_figures to be only a CompositeAnalysis option. I'm not sure why the lint test is failing—it's succeeding locally and the GitHub error message is on a file I didn't change.

[itoko]

LGTM, thanks!