Remember to implement arrays of 2D/3D/Lorentz vectors in Numba

PR #42 had to stop early because we need Numba to let us recognize np.ndarray subclasses inside of a Numba-compiled function. This feature is being implemented, but it is not finished or in any released version of Numba yet: numba/numba#6148.

When that is done, we can follow the procedure described in https://github.com/numba/numba/pull/6148#issuecomment-676541700 to:

• Override __getitem__ to return the appropriate VectorObject*DType when it would otherwise return a np.record.
• (Presumably, @DrTodd13's implementation would ensure that other slices maintain their class. If not, include that in the __getitem__ override. Currently, a VectorNumpy loses its class—becoming a plain np.ndarray—when it is range-sliced; this will have to be tested.)
• (It might also be necessary to override an array iterator.)
• Override __array_ufunc__ for the relevant ufuncs (np.absolute, np.add, np.equal, etc.).

That's it! Not nearly as much work as vector.backends.numba_object.

Start by reviewing PR #42 and the Numba feature in numba/numba#6148, then follow the procedure in its linked issue comment. Inspect VectorNumpy slicing and the relevant getitem and array_ufunc behavior, including range slices, record dtypes, and the listed ufuncs. Done means 2D, 3D, and Lorentz vector arrays work correctly inside Numba-compiled functions.