Improved numpy typing.

[AaronSartin]

ArrayLike is a general union. It drops a lot of the type information. I appreciated the extra info included in Annotated but a more specific type like npt.NDArray[np.float64] is often preferred for at least the return types. With no_convert enabled it is important for input types as well. Unfortunately this isn't a type that's easy to override with a type_caster because the type is generated in stubgen.py. I am running this sed command after generating my stubs to fix the typing:

"s/Annotated\[ArrayLike, dict(dtype='\([a-z0-9]*\)', /Annotated[npt.NDArray[np.\1], dict(dtype='\1', /g"

Is there a better solution already available? Would there be interest in updating stubgen.py to make this typing pattern the default behavior?

[AaronSartin]

The suggestion is closely tied with #983. See this comment for an example of the typing I would like to see.