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

[jpivarski]

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 numba/numba#6148 (comment) 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.

[henryiii]

vector/src/vector/_backends/numba_numpy.py

Line 11 in 1c4b81d

As mentioned in scikit-hep/vector#43, this capability is blocked by numba/numba#6148.

This PR was supplanted by numba/numba#7121, which was added in Numba 0.54, so this should be unblocked.