Enforce consistency between input numpy layout and expected Eigen object memory layout

[ManifoldFR]

Consider the following example:

template <typename MatType>
const Eigen::Ref<const MatType> asConstRef(Eigen::Ref<const MatType> mat) {
  doSomething(mat);
  return Eigen::Ref<const MatType>(mat);
}

// in BOOST_PYTHON_MODULE
  bp::def("asRef", asConstRef<Eigen::MatrixXd>);   // notice this is the *ColMajor* matrix
//

This will work:

A = np.eye(2, order="F")
aref = asConstRef(A)

assert np.array_equal(A, aref)
assert not aref.flags.owndata
assert not aref.flags.writeable

This will not:

A = np.eye(2, order="C")  # "C" is the numpy default
aref = asConstRef(A)

assert np.array_equal(A, aref)  # fails
assert not aref.flags.owndata  # ok
assert not aref.flags.writeable  # ok

After discussing with @jcarpent we came to the conclusion that the return statement in C++, due to how EigenAllocator<Eigen::Ref<cv MatType>, ...> is wired up, returns something pointing to the wrong memory - actually, to temporary C++ memory created by EigenAllocator and of which an Eigen::Ref is given to asConstRef. This temporary memory is a (col-major) Eigen::MatrixXd created to host a copy of the (row-major) Numpy input, and destroyed in the destructor of the rvalue_from_python_data handler before the function returns. What is inside the returned aref view is random memory.

This was discovered while working in PR #325.

Two possibilities:

• redefine the Eigen::Ref types to have dynamic stride as to allow mapping from column major to row major; rework the EigenAllocator
• disallow the above.