I'm getting circular imports trying to address ruff suggestions.

In numtype I've been using nested (hierarchical) .ruff.toml quite a bit. It avoids the main ruff config from becoming too monolithic when there are many per-file-ignores. For example in the src/numpy-stubs/@test dir there's a .ruff.toml with the contents:

extend = "../../../pyproject.toml"
line-length = 88

[lint]
extend-ignore = [
    "D",       # pydocstyle
    "ERA",     # eradicate
    "S",       # flake8-bandit
    "T",       # flake8-print
    "PLR2004", # pylint: magic-value-comparisons
]

Perhaps something like that could also help for numpy?

The amount of per file ignores remains limited for now. I suggest we revisit hierarchical ruff.toml files when adding more rules in the future.

I'm not sure how to use mypy_primer, there weren't too many typing-related changes here. I have rebased just in case.

Diff from mypy_primer, showing the effect of this PR on type check results on a corpus of open source code:

vision (https://github.com/pytorch/vision)
- torchvision/utils.py:271: error: Unused "type: ignore" comment  [unused-ignore]

https://github.com/pytorch/vision/blob/main/torchvision/utils.py#L271

🤷🏻

Rebased to fix conflicts.

needs another rebase

Rebased to fix conflicts.