Simplify shared union serializer logic

[sydney-runkle]

No description provided.

[codspeed-hq]

CodSpeed Performance Report

Merging #1538 will not alter performance

_{Comparing dh/wip-union-simplification (ee9dd3d) with main (061711f)}

Summary

✅ 155 untouched benchmarks

[davidhewitt]

Looking great! I think we can go a bit further for fun and profit 😄

[sydney-runkle]

Failing integration test is unrelated to this PR, should be resolved soon once we merge pydantic/pydantic#10825 into main on pydantic

[davidhewitt]

This looks like a solid improvement to me! 🎉

[MarkusSintonen]

@sydney-runkle did you try running CodSpeed in Pydantic side for this PR? I accidentally rebased my changes on top of pydantic-core master (instead of release tag) and pointed Pydantic side PR to point into the master based code (which included only this PR changes). CodSpeed showed this which is quite a regression:

❌ test_north_star_dump_json 31.7 ms 57.9 ms -45.26%

[davidhewitt]

I think the benchmarks we have here should have identified if this PR was a problem, but it's not impossible it could slip through.

[davidhewitt]

This PR should in theory have no functional change so definitely that order of magnitude seems incorrect.

[MarkusSintonen]

@davidhewitt / @sydney-runkle it is still showing the exactly same here: pydantic/pydantic#10845 🤔 Weird...

[sydney-runkle]

Huh, that is really odd. @davidhewitt, I might defer to you here - not sure why our changes here would degrade perf...

[davidhewitt]

@sydney-runkle this is the source of the perf regression; accidentally switched off tagged union serialization optimization in the JSON case.

[sydney-runkle]

😬 oops. Great find!