Add orth interface function

[sdrave]

This PR can be seen as an alternative approach to #2211. It adds an orth interface function that orthonormalizes a VectorArray by dispatching to various orthonormalization algorithms.

Main characteristics:

• If pad=False (which is the default), the spans of V and the computed basis are supposed to agree. Thus, len(basis) <= len(V). By setting pad=True, it is ensured that len(basis) == len(V), potentially by adding random vectors to the span.
• If hierarchical is True, the span of the first k vectors in V agrees with the span of the first k vectors in basis. I.e., we obtain a QR decomposition. Since orth returns the coefficients of V in the new basis as expected by lincomb, the returned array has to be transposed in order to get a QR decomposition.

The approach is based on the general observation that (more or less) everywhere in pyMOR, where qr (or gram_schmidt) is called, we actually want to orthonormalize a basis and do not care so much about the actual algorithm. Also, I think that orth fits better with the notion of a VectorArrray then QR decomposition. As such, I have replaced all calls to qr with orth. I wasn't always sure, of the resulting basis were needed. I see it as a feature of this approach that this has to be made explicit in orth calls. I also believe that a qr function would not really be needed in this case, or would act as an alias for orth.

This is meant as a basis for discussion. The implementation is ad-hoc and not fully tested. Argument names or their precise semantics, as well as the name of the function itself could be changed.