praefect: Read line-by-line in track-repositories

When parsing invalid input json.Decoder may go into an infinite loop and consume an unbounded amount of memory. Admins who accidentally pass a badly formatted file to track-repositories are at risk of accidentally triggering OOM kills as a result.

To mitigate this risk, let's read the input line-by-line and then attempt to unmarshall it, which will gracefully fail on invalid JSON. This does prevent us from checking for unexpected fields in the input, but so long as the expected fields are present extra data isn't a problem.