Change internal data storage format

[khaeru]

(Copied from the former doc/roadmap.rst.)

sdmx handles Observation as individual object instances. An alternative is to use pandas.DataFrame, xarray.DataSet, or other data structures internally.

See:

• sdmx/experimental.py for a partial mock-up of such code, and
• tests/test_experimental.py for tests.

Choosing either the current or experimental DataSet as a default should be based on detailed performance (memory and time) evaluation under a variety of use-cases.
The key question is: is there a performance limitation that this could solve?

To that end, note that the experimental DataSet involves three conversions:

1. a reader parses the XML or JSON source, creates Observation instances, and adds them using DataSet.add_obs()
2. experimental.DataSet.add_obs() populates a pd.DataFrame from these Observations, then discards them.
3. experimental.DataSet.obs() creates new Observation instances again.

For a fair comparison, the API between the readers and DataSet could be changed to eliminate the round trip in (1)/(2), but without sacrificing the data model consistency provided by pydantic on Observation instances.