Figure out all remaining ArrFunc slots on the dtype

[seberg]

Moves/partially moved (defined defintely on DType)

• Get/Set item
• copyswap(n) --> part of CastingImpl
• cast functions --> part of CastingImpl

Needs decision

• compare: also defined as ufunc, mostly needed for sorting.
  UFunc usage seems likely best, but could be better if
  UFunc can return n=0 specialized inner-loop
  implementation.

• nonzero(char *element, PyArrayObject *arr): Currently method (byte swaps
  if needed). Could also be a get_nonzero_func method?
  Also plausible as a ufunc?! (if we allow tagging on scalar specializations)

• fill (arange): Replace with better arange slot/method-like and
  move the weirder legacy behaviour through that?
  Problem: Either it needs to allocate, or getting the length needs to
  happen before somehow? A get_arange slot could also pass out the length
  although it is a bit absurd for such a simple function?

Deprecated or Unused

• clip # Internal usage deprecated
• fillwithscalar # Not used in NumPy

Should probably move to DType

• scanfunc, fromstr # As methods (get in the dtype?)

• dotfunc: Used for vector * (!matrix). Actually used to ensure that the
  32bit limit of standard BLAS is not hit.
  Needs decision whether to use get_dotfunc type slot, or just a single
  slot.

• sort, argsort: Needs new solution to avoid array of funcs.
  Should we change the sorting signature (aside from array->dtype change)?
  Could create a NPY_dt_get_sortfunction slot. That could return
  NotImplemented if the sortkind is not recognized. DType could be passed
  in any case (but dtype instance specialization would be possible).

Simply Remove

• take, putmask: Always NULL and never use within NumPy (done outside of this branch/work).

Move to GUFunc

• argmax/min: Should these be gufuncs simply? (although the gufunc may not be
  able to do the "ravel" logic, meaning it would be a wrapped gufunc).
  We can automatically do this for legacy types in the resolver.
  Since it can return a UFuncImpl which uses the dtypes function (if it
  exists). Only issue: That is typically only contiguously defined, but
  that is merely a performance issue, that only matters for types that
  probably do not even exist?

To Deprecate?

• scalarkind: Used only in PyArray_ScalarKind (which is public API).
  Nobody probably actually uses this (really). The public API
  function converts signed integers to unsigned integers.
  User dtypes could say they are integers/floats for scalar
  casting purposes.
  Scalarkind will probably just silently phase out with new dtypes and
  is never actually used by any extension dtypes.