Add 'random' method to VectorArrayInterface

[sdrave]

As discussed in #544, this adds a random method to VectorArrayInterface. It has a default implementation for all backends that implement from_numpy. The exact signature is

def random(self, count=1, distribution='uniform', random_state=None, seed=None,
           reserve=0, **kwargs):

where kwargs are additional parameters depending on distribution. The names of these parameters are chosen to agree with the parameters of the respective NumPy methods.

random_state and seed behave as for CubicParameterSpace.sample_randomly: either a (mutable) NumPy RandomState can be provided which is then used to sample random numbers, or a seed from which a new RandomState is initialized. If neither are provided a new random state is created using pymor.tools.new_random_state with a default seed.

An important consequence: repeated calls to the random method will always create the same random numbers when neither random_state nor seed are provided. This is probably unexpected behavior and I think it should be changed s.t. we have a global default RandomState which is initialized once with a fixed seed and which is always used unless the user provides a seed or RandomState. This should be handled in a separate PR, however.

I decided to always stick with NumPy RandomStates. In case a backend uses its own random generator, the NumPy RandomState should be used to create a new seed for the random generator of the backend. This way the interface of random stays generic.

[pmli]

@sdrave Looks good to me. Did you plan for this PR to also replace the from_numpy(np.random.randn(...)) stuff with the new method (in reductors/h2.py, reductors/sor_irka.py and algorithms/lradi.py)?

[sdrave]

@pmli, I think we should better do this in another PR after the RandomState issue has been resolved.