BUG: TYPE_setitem functions are broken in various ways #8902

Multiple bugs in one issue:

• LONGDOUBLE_setitem loses precision

  tiny = np.nextafter(np.longdouble(0), 1)
  assert tiny.itemsize > 8, "requires longdouble to test"
  tiny_arr1d = np.array([tiny])
  assert tiny_arr1d[0] != 0
  tiny_arr1d[0] = tiny_arr1d[0]
  assert tiny_arr1d[0] != 0
  tiny_arr1d[0,...] = tiny_arr1d[0,...]
  assert tiny_arr1d[0] != 0
  tiny_arr1d[0] = tiny_arr1d[0,...]  # forces a pass through `float()`, which discards precision
  assert tiny_arr1d[0] != 0  # AssertionError

• UNICODE_setitem goes via array.__str__() (causing Unicode scalar problem  #5049 on python 2.7):

  from contextlib import contextmanager
  from numpy.core.numeric import set_string_function
  
  @contextmanager
  def inject_str(s):
      set_string_function(lambda x: s, repr=False)
      try:
          yield
      finally:
          set_string_function(None, repr=False)
  
  q = np.array([u'abc'])
  q0 = np.array(u'def')
  with inject_str('WAT'):
      q[0] = q0
  assert q[0] == q0  # AssertionError: it's WAT instead

• UNICODE_setitem and STRING_setitem behave differently when __len__ fails

  class broken:
      def __len__(self):
          raise ValueError('no length')
      def __repr__(self):
          return 'OK'
      def __getitem__(self, i):
          return 42
  
  u = np.array([u'test'])
  u[0] = broken()
  assert u[0] == 'OK'
  
  b = np.array([b'test'])
  b[0] = broken()
  # ValueError: cannot set an array element with a sequence

  Arguably the correct thing to do is just propagate the exception

• UNICODE_setitem and STRING_setitem have a weird special cases for empty sequences:

  s = np.array(["test"])
  s[0] = []
  assert s[0] == "[]"
  s[0] = [1]
  # ValueError: setting an array element with a sequence