[ENH] speculative design for raw data container DataSet

[fkiraly]

This PR contains a speculative design for a data container TimeSeries.

This follows the idea to split the current TimeSeriesDataSet in two parts:

• a "raw data set" object, representing a collection of time series or a single time series - contained in this PR as docstring
• a "resampled data set" which is based on a raw data set and additional resampling instructions - not contained in this PR

The idea of a split was discussed in #1736, but also already occurs in the comments of @jdb78 in the current TimeSeriesDataSet.

• Choices for input and output fields are based roughly on the DSIPTS design, and splitting TimeSeriesDataSet in two sets of concerns, one raw data specific, one resampling/batching specific.
• names are chosen that they are descriptive and short, making usage typing efficient.

The pandas dependency can be removed by changing the input while not changing the output format in this first raw data layer, inheriting from a common base.

[fkiraly]

Some self-review:

• the output format seems a bit complex - can we reduce this to less tensors and still keep handling of categorical and future data?
• where exactly is the separation into past and future data needed? Do we already need this in the raw data layer, or can we keep it as metadata? Do we need to allow for data_future to be passed?