BUG: linalg: norm fails on longdouble, signed int

This fixes the following bug with longdouble:

>>> x = np.arange(10, dtype=np.longdouble)
>>> np.linalg.norm(x, ord=3)
Traceback (most recent call last):
  File "<ipython-input-5-7ee53a8ac142>", line 1, in <module>
    np.linalg.norm(x, ord=3)
  File "/tmp/v/lib/python2.7/site-packages/numpy/linalg/linalg.py", line 2090, in norm
    return add.reduce(absx**ord, axis=axis)**(1.0/ord)
UnboundLocalError: local variable 'absx' referenced before assignment

As well as the handling of minimal values for signed integers:

>>> x = np.array([-2**31], dtype=np.int32)
>>> np.linalg.norm(x, ord=3)
/tmp/v/lib/python2.7/site-packages/numpy/linalg/linalg.py:2090: RuntimeWarning: invalid value encountered in double_scalars
  return add.reduce(absx**ord, axis=axis)**(1.0/ord)
nan

0 files changed, 0 insertions, 0 deletions