dtype promotion rules are not quite right for element-wise ufuncs

[shoyer]

Many ufuncs actually always converge this argument to float, e.g., consider:

In [13]: values = np.random.randint(10, size=(4, 6))

In [14]: data = da.from_array(values, blockshape=(2, 2))

In [15]: data.dtype
Out[15]: dtype('int64')

In [16]: da.sin(data).dtype
Out[16]: dtype('int64')

In [17]: da.sin(data).compute()
Out[17]:
array([[ 0.14112001,  0.14112001,  0.14112001,  0.90929743,  0.41211849,
         0.        ],
       [ 0.6569866 ,  0.14112001, -0.2794155 , -0.2794155 ,  0.14112001,
         0.90929743],
       [ 0.        ,  0.90929743,  0.84147098, -0.7568025 ,  0.98935825,
         0.90929743],
       [-0.95892427,  0.90929743,  0.14112001,  0.14112001,  0.        ,
         0.        ]])