Build merged_draws.parquet and merged_summaries.parquet

[kgostic]

@zsusswein will review

[natemcintosh]

So, going off the output structure docs, the raw output will have form

az://rt-output/
├── job_<job_id>/
│   ├── raw_samples/
│   │   ├── samples_<task_id>.parquet
│   ├── summarized_quantiles/
│   │   ├── summarized_<task_id>.parquet
│   ├── diagnostics/
│   │   ├── diagnostics_<task_id>.parquet
│   ├── tasks/
│   │   ├── task_<task_id>/
│   │   │   ├── model.rds
│   │   │   ├── metadata.json
│   │   │   ├── stdout.log
│   │   │   └── stderr.log
│   ├── job_metadata.json

• If re-runs end up in the same job, that makes created merged_... files a little simpler, as they can go in the same job folder.
• If re-runs end up in a different job, then we would put the merged... files in the production database?

I guess I'm slightly unclear on the difference between these merged_... files and the "production database".

[natemcintosh]

Assuming the section in this PR updating the README on "outputs" is up to date, I think I understand how this all fits together.