JSON encoding engine instrumentation

[jonmmease]

This PR should not be merged!

This PR is instrumented for testing the new JSON encoding engine being introduced in #2955 (comment). Idea from @nicolaskruchten in #2955 (comment).

With this branch, every call to to_plotly_json performs JSON encoding with all three engines, then decodes the result and checks that identical dictionaries are produced. When they are not identical, and exception is raised.

Locally, nearly all of the documentation examples work. The two exceptions are imshow and ml-tsne-umap-projections. Both of these fail due to difference in how the orjson encoding handles floating-point numbers with precision less than 64 bits.

When encoding a float32 numpy array, the legacy and json encoders will convert this to a list of regular Python 64-bit floats before encoding them. This results extra digits in the JSON representation. In contrast, the orjson encoder will only write out the number of digits corresponding to the original 32-bit values.

[nicolaskruchten]

awesome! To get a bit more mileage out of our extant tests which don't call fig.show() maybe add a JSON call to the end of px._core.make_figure() or something? Maybe consider adding it to all the update_* methods so that we check a bunch of intermediate representations as well? Super-slow, I know, but doing it just a couple of times would do ;)

[jonmmease]

Yeah, those are good ideas. I'll add that sometime over the next couple of days

[jonmmease]

@nicolaskruchten, I added the extra instrumentation points in be81b16 and everything looks the same as far as I can tell. See what you think!

[nicolaskruchten]

OK, are you just running these locally or...? CI is failing because orjson isn't installed so I assume your goal wasn't to get CI to go green here, right?

[jonmmease]

Yeah, I was running locally. And on CI I was mostly looking at the docs. I had the orjson stuff working on CI, but maybe that's in the other PR. I'll take a look.

[jonmmease]

Merged in #3022, and a couple CI fixes and now all existing tests are passing with the instrumented version.