Fix long import times due to processing all pydantic subclasses

[jaredoconnell]

Summary

Other libraries that use pydantic could end up being processed by GuideLLM, significantly increasing the number of classes processed, noticeably increasing import times for GuideLLM.

Also, in my testing, it makes it so the startup times are less than half of what they were before.

Details

• Rather than just searching every Pydantic subclass, it only searches sub-classes that we created. So no other libraries will add to the processing burden.
• On my mac, within the Python virtual environment, this change resulted in a change from around 27.5 seconds before the benchmark starts to 6.5 seconds. This is due to how long it spent processing the classes on my local env. It has been super slow since this was originally added in the refactor.
• The only classes that use this functionality are BenchmarkConfig, BenchmarkGenerativeTextArgs, and Profile,

Test Plan

• Run benchmarks as normal, and confirm that the options that depend on registry mix in work.

Related Issues

• Resolves import guidellm can take a very long time in some environments #688

---

• "I certify that all code in this PR is my own, except as noted below."

Use of AI

• Includes AI-assisted code completion
• Includes code generated by an AI application
• Includes AI-generated tests (NOTE: AI written tests should have a docstring that includes ## WRITTEN BY AI ##)

[dbutenhof]

Excellent! But, as Webb might say, just a few nits ...

[dbutenhof]

"for all GuideLLM Pydantic models ..." ?? Yeah, "that reference this" covers that base, but it's a bit less direct.

[jaredoconnell]

What type of change would you suggest?

[dbutenhof]

Ha -- you might consider modifying the PR title: it's an incomplete sentence, and it displays as "Fix long import times due to #689", which just looks odd!