allow model to map a linear set of weights

Currently, the benchmark imposes a certain linear mapping on the model (e.g. PLS/Ridge).
We should give modelers the freedom to use their own method of finding a linear translation.
The API could look something like this:
```
class BrainModel:
...
def find_linear_weights(self, self_recordings, target_recordings)
-> [DataAssembly(self_neuroid x target_neuroid), DataAssembly(target_neuroid)] # linear matrix, bias
```
In this setting, cross-validation and thus available data would still be in the hand of the benchmark but the model can employ its own method for finding a linear translation. We would still provide standard mapping procedures (e.g. PLS / mask regression).