Add a wrapper for the MUMPS null space algorithm

Closes #7

See merge request kwant/mumpy!5

Closes #4

See merge request kwant/mumpy!7

Add tests for all datatypes.

Closes #2

See merge request kwant/mumpy!4

Using a large, dense and completely random matrix with a sparse
solver takes a long time, but this test is only testing whether
MUMPSContext correctly increases the memory limit when it detects
that MUMPS has run out. This means we can choose a matrix that
is sparse and easy to factorize, as long as the memory requirements
are large.

We're just checking that everything is wired up correctly, and
applying the zero-one-infinity rule tells us that we don't need
to test many different matrix sizes for this. Prior to this change
the tests took ~10 seconds to run, now they take <1 second.

There is no need to take the difference of the two arrays, nor
to take the absolute value or max (assert_array_almost_equal already
does this for us). Similarly, there is no need to compare the real
and imaginary parts separately.

Use more concise language, and make it clear where the precision
limits originate from. Also, use proper comments as opposed to
string literals.

Kelvin Loh authored 6 years ago and Joseph Weston committed 6 years ago

Change the criterion for array equal assertion
to the infinity norm of the residual.
Remove the kwant dependency of the test script.

This had been forgotten when bringing over Kwant's setup.py

Closes #10.

Kwant's setup.py handles config file parsing and a lot of edge
cases related to Cython, so we keep it around even if it is a bit
long.

Closes #3.

Closes #1.

Closes #9.

Setup fixes

See merge request kwant/mumpy!1

Mumps splits single precision, double precision etc. versions
into separate libraries, so we should link against them all

This is completely irrelevant for Mumps and is a hangover from copying
code from Kwant.