Add benchmarks for ExtractSubMatrix + CopySubMatrix

[fingolfin]

These functions are part of the MatrixObj AP, but currently they are basically always slower than their list arithmetic counter parts involving the m{}{} syntax, even in situations where those APIs were supposed to be superior.

I believe this can be reversed by providing tuned methods for these functions (at least for the built-in plain and compressed matrices). In order to tackler this systematically, we need robust benchmarks to measure the effect of any changes.

This patch adds a first version of such benchmarks. Of course what we really need are the improvements, but now at least we have a baseline.

[fingolfin]

Same should be done for ExtractSubVector and CopySubVector and we should add (eager?) methods for those handling classic "plain" vectors efficiently, so that code can use them without reservations.

It might also be useful to consider allowing the mat{list1}{list2} syntax apply to general MatrixObj, and creating new matrices efficiently. This may (???) require some kernel work, but even if that shouldn't be too hard. But it would mean more existing code could stay as it is and still support general MatrixObj matrices.