This avoids the memory allocation. It is very strict in checking that
the types are correct, but since it is intended as an optimisation, it
should only be used when the user knows what they are doing.

Using the toy example in the test as a benchmark, I clocked 63ms with
an `out` parameter, versus 91ms without. Here is the full
benchmark (requires ipython)::

    import numpy as np
    np.random.seed(22)
    f = np.random.rand(1024*128, 16)
    v = np.random.rand(16,32)
    r = np.empty((1024*128, 32))
    _ip.magic("timeit np.dot(f,v,r)")
    _ip.magic("timeit np.dot(f,v)")

(Pointed out by pv on github)

This allows `np.dot(a, b, out=r)`.

Suggested by pv.

…ture in PyArray_MatrixProduct2

This indicates that the function takes 3 arguments