git-ref: packed-refs backend for files store MVP

[Byron]

As a follow-up of #90

• research - it seems clear how it works and how it should be implemented competitively. They are zero copy, too :D.
• parse header line
• parse records
• lookup records

Research

• mmap can be a problem on windows - rather read it all into memory and use reftable if that's too much at some point.
• big files of 150k packed-refs (see this script to generate it
• git essentially parses each line on the fly and calls the owning structure rightfully 'snapshot'. Most if not all operations are done directly on these. The faster newlines are detected the better, validation happens too.
• the find implementation is a binary search over snapshot records.
• Not all refs are packed, symbolic refs for one are missing

[Byron]

Closing in favor of #104