Add support for 'arraylike' objects as JSON arrays

[calgray]

JMESPath.py is limited in that only the dict and list derived containers returned by the built-in json library are supported in the object hierarchy due to the use of isinstance. A very notable arraylike instance that does not derive directly from these containers is a numpy.ndarray which can be deserialized using the JSON-like msgpack library with msgpack_numpy.

This changeset aims to add support for arraylike (list, tuple and numpy.ndarray) containers in place of parsed JSON arrays and without adding any dependency on the numpy library. This is done using the documented numpy array interface protocol of which many more arraylike libraries adhere to such as xarray, dask, astropy and cupy.

(pandas.Series is also arraylike but limited to 1D as multidimensional Series isn't an intended use case and has slicing issues)

[edgarrmondragon]

Worth linking to https://numpy.org/doc/stable/user/basics.interoperability.html#the-array-method in a comment?

Also, is it guaranteed that an object with __array__ will always have a shape attribute defined?

[calgray]

The numpy docs describes that __array__ method if it exists should always return a np.ndarray instance (ideally with zero copy) of which that type always has a shape attribute (tested with dask and astropy).

__array_interface__ is a bit more array library agnostic and explicitly documented to require shape, but __array__ is already being used so this project doesn't need to explicitly import numpy to perform np.array(value.__array_interface__, copy=False).

[calgray]

Oh, I see what you mean. Updated to:
return hasattr(arg, "__array__") and arg.__array__.shape != ()