Add specification for explicitly casting an array to a specified dtype

[kgryte]

This PR

• specifies an astype API for explicitly casting an array to a specified dtype.
• adds a functional API rather than an array object method following conventions set forth in this specification to maintain a minimal array object API.
• uses the astype name to mirror the NumPy array object method x.astype().
• includes a copy kwarg to allow returning the input array iff copy is False and the specified dtype matches the data type of the input array. Every library except TensorFlow supports this kwarg. During the consortium meeting on 21 October 2021, TF supported including the copy kwarg in the specification.
• updates conversion guidance for asarray to only require support for conversions following type promotion rules described in this specification. With the addition of astype, overloading asarray with comprehensive casting semantics for input arrays is not necessary.
• resolves Cross-kind casting in asarray() #264.

Prior Art

• TF cast: only supports providing a dtype argument.
• NumPy x.astype: supports order, casting, and copy kwargs.
• CuPy x.astype: supports order and copy kwargs.
• Dask x.astype: supports casting and copy kwargs.
• Torch to: supports copy and device kwargs.
• MXNet x.astype: supports copy kwarg.

[kgryte]

This change was discussed and agreed upon during the consortium meeting held on 21 October 2021. Merging...