The strings in error messages were left untouched

Reloading currently causes problems because global classes defined in
numpy/__init__.py change their identity (a is b) after reload. The
solution taken here is to move those classes to a new non-reloadable
module numpy/_globals and import them into numpy from there.

Classes moved are ModuleDeprecationWarning, VisibleDeprecationWarning,
and _NoValue.

Closes #7844.

There seems to be little use in reloading numpy as any changed modules
that are imported into numpy would need to be reloaded first in order to
see any changes. Furthermore, reloading causes problems as global
classes defined in numpy/__init__.py change their identity. Hence we
raise a RuntimeError when an attempt to reload numpy is made.

Closes #7844.

Give hook to allow platform-specific installs to modify the
initialization of numpy.  Particular use-case is to allow check for SSE2
on Windows when shipping with ATLAS wheel.

Our test-runner's raise_warning mode traditionally has varied depending
on whether we have a development or release version of numpy: for
development versions we raise on warnings, and for release versions we
don't. This is all very sensible... *if* you're running numpy's test
suite. But our test-runner is also used by other packages like scipy,
and it doesn't make sense for scipy's raise_warning mode to vary
depending on whether *numpy* is a development or release version. (It
should vary depending on whether the scipy-under-test is a development
or release version.) So this commit moves the numpy-version-dependent
raise_warning logic out of the generic NoseTester class and into
numpy-specific code.

(See scipy/scipy#5609 for more discussion.)

See gh-432 for details.  Motivation for adding this now is the discussion
on gh-5343.

Add _NoValue class at top level to make it possible to detect when
non-default values got passed to a keyword argument, as in:

    def func(a, b=np._NoValue):
        if b is not np._NoValue:
            warnings.warn("Argument b is deprecated",
                          DeprecationWarning)

interpreter is misspelled

This function is commonly confused with numpy.linalg.matrix_rank
and exists itself only for history reasons.

Closes gh-4616