This fixes the distutils built from an sdist (e.g. under tox),
without including _build_utils in binary distributions or the installed numpy.

Allows access to internal functions for the file.

Add our python_xerbla to the multiarray sources. That function is
needed for all modules that link to the ATLAS 3.10 libraries, which
are now all located in two files, libsatlas and libtatlas.

Also make the test for xerbla linkage work better. If xerbla is not
linked the test will be skipped with a message.

Closes #5362.

MAINT: merge scalarmathmodule into umath module

minimum requirement is python2.6, this allows removing a couple 2.3 and
2.4 checks.

There is no good reason it needs to be its own module.
Also removes the undocumented pymath alter/restore functions.

For cases like convolve/correlate with small kernel calling out to BLAS
dot is very slow as it goes through many wrappers and BLAS is typically
optimized for very large data.
So add a template utility file and add a special function for small dot
products.
2-4 times faster than openblas dot for kernel sizes < 11.

Remove vdot from _dotblas and implement it in multiarray.
Remove the files in core/blasdot as they are no longer needed.
Fix tests and build to reflect the changes.

The vdot function is now a bit faster in the common case as ravel
is used instead of flatten. There is also no need to conjugate the
files for clongdouble.

Move the dotblas_matrixproduct function from the _dotblas.c file to a
new cblasfuncs.c file in multiarray and rename it cblas_matrixproduct.
Modify it so that it can called directly from PyArray_MatrixProduct2 and
do so.  Fix up numeric.py and core/__init__.py to reflect these changes.