Try to fully resolve the NumPy format string issues.

[jpivarski]

• util::dtype enum for all the recognized primitives.
• functions for converting between (format, itemsize) and util::dtype.
• util::dtype npdtype as a third type parameter in NumpyArray and NumpyForm, part of the constructor.
• Replace uses of format with uses of npdtype without changing behavior.
• Prefer npdtype when converting NumpyArray to and from np.ndarray.
• Use the same mechanism for IndexOf<T>.
• Extend NumpyArray implementations to all util::dtype values (except NOT_PRIMITIVE).
• Extend reducers as well.
• Revisit the linked issues to see if they've been fixed:
  • Make NumpyArray's format checks insensitive to '=' (or '<' if little-endian, '>' if big-endian) at the beginning of the string #313
  • Layout indexes don't support slices #333
  • Numpy dtype of Index64 #335
• NumpyForm JSON string representations should cover all of the util::dtype values (except NOT_PRIMITIVE).
• We need a function that turns NumPy dtypes into NumpyForms. The JSON trick is dumb.

[jpivarski]

Initial survey: pybind/pybind11#1908 (comment)