Speed create_import_structure up with os.scandir()

[Rocketknight1]

create_import_structure_from_path does some redundant os calls, so I'm experimenting with changes to see if we can speed up loading a lot.

Related to #44246

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[Rocketknight1]

cc @LysandreJik while you're working on this! I don't think it addresses the contributor's issue but it might be some free speedup regardless.

The tl;dr is that the existing code does two wasteful things. Firstly, it doesn't skip convert_*.py files in the module dir, which are never relevant. Those files are excluded in release versions but are still present if people install from main.

Secondly, the old approach of os.listdir(module_path) followed by os.path.isdir() requires multiple system calls per model, but it's much faster to use os.scandir because the OS actually returns the information we need in the directory listing call, so we avoid thousands of system stat calls this way. The speedup may not be huge, but there's no downside, so it's probably worth merging!

[Rocketknight1]

(Failing test is unrelated and goes on my list of flaky tests to fix)

[LysandreJik]

Thank you @Rocketknight1 !