Ensure that np.unique works on data types with special comparators.

Revert "BUG: ticket #2063, make unique return consistent index."

This reverts commit 74b9f5eef8fac643bf9012dbb2ac6b4b19f46892.

Conflicts:
	numpy/core/tests/test_regression.py

This commit introduced a regression in `np.unique` for certain data types:

  >>> A = np.array([[1, 2], [1, 3], [1, 2]], dtype='i')
  >>> B, I, J = np.unique(A.view([('', A.dtype)]*A.shape[1]), True, True)
  Traceback (most recent call last):
    File ".../numpy/lib/arraysetops.py", line 178, in unique
      perm = ar.argsort(kind='mergesort')
  TypeError: requested sort not available for type

See #1664.
Thanks to Matthew Brett for suggesting this improved detection of PPC.

Conflicts:

	numpy/core/tests/test_umath.py

[Back-ported from master by njs, with an additional tweak to
memmap.__del__ to avoid unhandled-exception warnings.]

This test was reported to be failing once for 1.6.2rc1, test precision may be
set slightly too agressive.

The --cover-inclusive argument means that coverage.py tries to include every
single .py file in the source tree in the coverage report. This leads to test
errors, because it tries to import files like setupscons.py (which will of
course directly fail for anyone not having numscons installed).

(backport of bfaaefe5)

Backport numpy.random fix for MSVC.

Backport numpy.poly bug fixes and documentation improvements.

Backport f2py fixes and improvements.