Improve BenchmarkDataset (and Benchmark) for more flexibility

[ben-jy]

The BenchmarkDataset class (and the Benchmark class) should be improved to cover a wider range of use cases. To do :

• Add parameter to define a batch size, as it depends strongly on the samples dataset dimension
• Add parameters to choose which components are used as input respectively as target (currently, all components are used as input as well as target), reflecting real-world applications.
• Make the given time series instantiable dynamically, limiting memory usage typically when a lot of datasets are benchmarked
• Add parameter to define a pre-processing pipeline. For now, it can be a simple lambda.