Fix labelled data counts not covering all selected train IDs

[philsmt]

When building tests for exdf-tools, I noticed a difference in behaviour between labelled and unlabelled KeyData.data_counts(). Whereas the latter always covers the selected train IDs, the former only contains only those described by data chunks for this key.

This situation can arise if a source is missing entirely in one or more sequences as opposed to being present with no data counts. In practice, this can happen e.g. when:

• A device goes down during a run, causing the aggregator to close the current sequence and start a new one
• Combining two runs with a source only present in one of them
• A source being dropped from sequence due to train selection in exdf-reduce (what triggered this observation)

While this is somewhat recoverable given the returned series is labelled, I would argue our APIs generally cover all selected train IDs and hence I would fix this for consistency.