issues/PRs that go towards combining datasets / responses go here

1. (DONE) Combine gen datas on forward_model_ok and in responses.json
2. (DONE) Use parquet format (prefer writing it w/ polars) for responses and observations
3. (DONE) Move observations parsing to configs
4. (Partially done) Make response configs generic (there are still some logic specific to response types in dark storage)
5. Add more performance tests to see performance improvements/regressions: 8a944a9 , 68c2de9 , 233bee2
6. One column per response key(?) - look into potential performance gains, should be faster due to not having to scan&filter per response key. (tests were green before join memory fix, join logic needs to be revised, but otherwise done) Try store response key as columns #9360
7. Combine datasets (not done, but also not much missing) Combine responses across realizations #9338
   Statemap: Should be much faster than before, due to fe9926b , 8848f37 .
   Could make it faster by saving statemap to file whenever we update the cache.