[MrDMD] Enable per-level (and per-leaf) choice of DMD implementation

[fandreuz]

There are some use cases (see #251) in which it would be convenient to select different kinds of DMDs according to the level inside MrDMD, or to change some parameter depending on the depth.

In this PR I added the method _dmd_builder, which provides DMD-builder functions and generalizes the current approach for the construction of DMD instances. After this PR the parameter dmd of the constructor of MrDMD accepts values of type list, tuple, function and DMDBase. See the documentation of MrDMD._dmd_builder for more details and examples.

Example

We use a different kind of DMD if we are near the middle part of the time window:

>>> def my_dmds(level, leaf):
...     level_size = pow(2,level)
...     distance_from_middle = abs(leaf - level_size // 2)
...     # we choose 2 as a random threshold
...     if distance_from_middle < 2:
...         return HankelDMD(d=5)
...     else:
...         return DMD(svd_rank=3)
>>> MrDMD(dmd=my_dmds, max_level=5).fit(X)