different dtype when calculating division compared with numpy

[acphile]

For example:
in mxnet.numpy
   for two ndarray a,b with dtype='int64', the dtype of result c = a / b is 'float32'
in numpy
   the dtype is 'float64'
Generally, when calculating division, mxnet.numpy often uses float32 while numpy uses float64.
Using float32 will often lead to the loss of floating point precision.

b=mx.np.array(1, dtype='int32')
e=b/3
e.item()
>>> 0.3333333432674408

[sxjscience]

This is to cope with the DL scenario. The default dtype in DeepNumpy is float32. See also #18060 and the current PR: #18251

[leezu]

Any diverging behaviour with numpy should either be fixed or clearly documented in a central reference location. Let's track either solution here

[sxjscience]

@leezu I suggest removing the Bug label because this is not a bug and won't be fixed. We could add the Document label instead.

[yzhliu]

@JiangZhaoh will be helping to document.

[yzhliu]

closed by #18399