Update default behavior of derivative of linalg.norm and abs at x=0

[rsokl]

(I also smuggled in a fix for mygrad.arange that normalizes its interface with that of numpy.arange

Now both mygrad.linalg.norm and mygrad.absolute, by default, will return 0 wherever their derivatives would typically not defined (and would previously have returned nan).
Both of these functions can be passed nan_to_num=False to enable the previous, more-rigorous, behavior.

Before:

>>> import mygrad as mg
>>> x = mg.tensor([-2.0, 0.0, 2.0])

>>> mg.absolute(x).backward()
>>> x.grad
array([-1., nan,  1.])

After:

>>> import mygrad as mg
>>> x = mg.tensor([-2.0, 0.0, 2.0])

>>> mg.absolute(x).backward()
>>> x.grad
array([-1., 0.,  1.])

>>> mg.absolute(x, nan_to_num=False).backward()
>>> x.grad
array([-1., nan,  1.])