Fix compilation for both ColMajor/RowMajor + provide a define.

[bchretien]

Users can use the CMake option "STORAGE_ORDER" to change the storage order used by RobOptim. It defaults to ColMajor (Eigen's default), and RowMajor can be used for hypothetical increased performance in some scenarios (e.g. when mostly using impl_gradient to fill the
Jacobian matrices).

Details

In RobOptim, gradient_t describes a row of the Jacobian matrix, but until now, it was defined as a column vector. A new rowVector_t is now available (vector_t is still a column vector), and in the dense case, gradient_t is just a rowVector_t. The sparse case is a bit different, since vector_t/rowVector_t are dense vectors but gradient_t is a sparse vector (now a RowMajor sparse vector for consistency with the dense traits).

Also, I encountered the following Eigen bug: http://eigen.tuxfamily.org/bz/show_bug.cgi?id=980
This will be fixed in the next Eigen release.

[coveralls]

Coverage decreased (-0.01%) to 92.41% when pulling 5597f65 on bchretien:master into 039e21f on roboptim:master.

[bchretien]

Note: plugins need to be tested before merging this.

[bchretien]

Things seem to be working as expected. The Ipopt plugin and the Python bindings needed a quick fix.

[coveralls]

Coverage decreased (-0.01%) to 92.41% when pulling a4dcac7 on bchretien:master into 039e21f on roboptim:master.

[bchretien]

I added an extra type, derivative_t, which is merely a typedef to a column vector. It should be used instead of gradient_t (row vector) for NTimesDerivableFunction (i.e. in roboptim-trajectory).