[x] Factor out JSON parsing and validation logic from import storages into a function that can be hot-swapped in LSO/LSE
[x] add it to settings as an app-specific variable
[x] replace stdlib parsing with pyarrow parsing to handle JSONL as well as JSON
[x] test coverage for JSONL
[x] feature flag
[x] unskip multitask import tests in LSO, to prepare for testing different behavior in LSE and LSO

Implementation notes:

• had to move computation of row_index and row_group into get_data so that load_tasks_json, which will be swapped in LSO/LSE, will have the correct type in all cases - makes for a pretty large diff
• simplified error messages significantly since the relevant logic is shared/deduped
• had to stick with stdlib json parsing instead of removing it entirely:
  • in the case of single dict, pyarrow can't distinguish between this and a one-line JSONL file, so need stdlib json to set row_index=None
  • in the case of top-level list, pyarrow can't parse this at all since it assumes each line is an entry in a table. There is an open issue for this.
  • consider peeking at file extensions to avoid double-parsing (for clarity, not speed) - see TODO comment