DOC: Describe return_counts keyword in np.unique docstring

Current loadtxt with `comments=None` considers the string `'None'` as a
comment symbol. Fixed by making split_line method check if comments is
None.

Closes #5155.

BUG: copy inherited masks in MaskedArray.__array_finalize__

* Previous expected behavior was that the gradient is computed using central
  differences in the interior and first differences at the boundaries.

* gradient was updated in v1.9.0 so that second-order accurate calculations are
  done at the boundaries, but this breaks expected behavior with old code, so
  `edge_order` keyword (Default: 1) is added to specify whether first or second
  order calculations at the boundaries should be used.

* Since the second argument is *varargs, in order to maintain compatibility
  with old code and compatibility with python 2.6 & 2.7, **kwargs is used, and
  all kwargs that are not `edge_order` raise an error, listing the offending
  kwargs.

* Tests and documentation updated to reflect this change.

* Added `.. versionadded:: 1.9.1` to the new optional kwarg `edge_order`
  documentation, and specified supported `edge_order` kwarg values.

* Clarified documentation for `varargs`.

* Made indentation in docstring consistent with other docstring styles.

* Examples corrected

Previously, operations which created a new masked array from an old
masked array -- e.g., np.empty_like -- would tend to result in the new
and old arrays sharing the same .mask attribute. This leads to
horrible brokenness in which writes to one array affect the other. In
particular this was responsible for part of the brokenness that
@jenshnielsen reported in gh-5184 in which np.gradient on masked
arrays would modify the original array's mask.

This fixes the worst part of the issues addressed in gh-3404, though
there's still an argument that we ought to deprecate the mask-copying
behaviour entirely so that empty_like returns an array with an empty
mask. That can wait until we find someone who cares though.

I also applied a small speedup to np.gradient (avoiding one copy);
previously this inefficiency was masking (heh) some of the problems
with masked arrays, so removing it is both an optimization and makes
it easier to test that things are working now.

BUG: fix nanmedian on arrays containing inf

BUG: Fix writing of intrinsic long integers in python2 npy files.