Compaction reliability

### User Story

As a user of Sleeper, I want compaction jobs to run deterministically and reliably, so that I know my data will be processed and won't be lost.

### Description / Background

We'd like to ensure that:
- Data scheduled into a compaction job will always be written to a new file.
- The new file replaces the old file atomically, such that no data is lost or duplicated.
- If data cannot be processed, it should be transparent to the user why this was the case, and what they can do about it.
- If a compaction job cannot be processed for some temporary reason, the underlying data will still be correctly processed.

### Technical Notes / Implementation Details

See sub-issues. We may make notes here on potential issues that have not yet been raised.