This pull requests adds annotations for ndarray and generic arithmetic operations:

• __add__
• __sub__
• __mul__
• __truediv__
• __floordiv__
• __pow__

A few notes about the PR:

While adding annotations to timedelta64 I encountered numerous issue due to its incompatibility with its signedinteger baseclass:

• No support for __index__
• No support for __pow__
• Does not support operations with complex numbers
• Most operations do not support floats

For this reason its baseclass has been changed from signedinteger to generic. While not strictly true during runtime, the problems with timedelta64 inheriting from signedinteger have been noted before (#10685) and this change would be inline with the fix proposed in abovementioned issue.

Secondly, one obstacle that hasn't been dealt with is how the handle generics precision,
e.g. how to annotate int64(...) + float32(...) -> float64(...) without adding massive amounts of overloads.
I might have an idea about to do deal with this issue, but I'd prefer to save it for a future PR due to some necessary compromises.

In any case, for the time being the return types are generally limited to types with ambiguous precision (signedinteger, floating, complexfloating, etc...)

import numpy as np

i8 = np.int64(...)
f4 = np.float32(...)
f4 + i8  # note: Revealed type is 'numpy.floating'

Finally, in order the deal with the repetitive and @overload-prone function signatures, this pull request makes extensive use of callback protocols. Callback protocols are effectively more fancy version of typing.Callable and, therefore , can be assigned and reused.