By avoiding writing metadata that differs but doesn't change the data,
we get a higher hitrate on the hash-checking when combining.

Use --debug=process to include these details for debugging.

When generating many parallel data files, often some data files will be exact
copies of each other.  Checking the hashes, we can avoid combining the
duplicates, speeding the process.

On a coverage.py metacov, we had 651 duplicates out of 2189 files (29%).
The time to combine was reduced by 17%.