Group observations by response type

[yngve-sk]

Towards #7279 Adds one dataset per response type, rather than one per observation key. Should speed up things some/quite a bit, especially for cases where the amount of observations by far exceeds the number of responses.