Explicitly support numpy arrays in foreach_get and foreach_set

[Andrej730]

Though we did included typing.Buffer as possible argument in #257 those methods still cannot accept numpy arrays.

typing.Buffer requires __buffer__ method to be implemented on class and in Python 3.11 though it wasn't exposed to Python (e.g. np.eye(3).__buffer__ will result in AttributeError: 'numpy.ndarray' object has no attribute '__buffer__') it was implemented internally in C-extension and buffer copy does work.

array.array had a similar thing and what they did, they virtually implemented __buffer__ method in the stub file (python/typeshed#10225) - basically just added to to .pyi though accessing it will result in AttributeError too. Though it might not be entirely correct, it probably what users are expecting from Buffer protocol.

I've created an issue in numpy (numpy/numpy#26783) and submitted a PR (numpy/numpy#26784) to do the similar thing but it wasn't accepted to avoid confusion with virtual method floating around.

I was thinking, can we extend seq argument in foreach_get / foreach_set with npt.NDArray explicitly? Numpy seems to be the most convenient container for those methods as it does support very efficient buffer copy - https://b3d.interplanety.org/en/optimizing-the-speed-of-data-access-using-foreach_/

Test example:

import bpy
import numpy as np

mesh = bpy.data.meshes[0]
verts_ = []
mesh.vertices.foreach_get("co", verts_)
verts = np.empty(len(mesh.vertices) * 3, dtype="f")
# Argument of type "NDArray[Any]" cannot be assigned to parameter "seq" of type "MutableSequence[bool] | MutableSequence[int] | MutableSequence[float] | Buffer" in function "foreach_get"
mesh.vertices.foreach_get("co", verts)

[JonathanPlasse]

I guess it is ok, numpy is always installed in Blender.
People will have to install numpy as a dev dependency for Pyright not to complain about unknown something when in strict mode.