Leverage S3 SnowBall Archive to speedup transfer

This seems to be complex to get sustainable performance when copying small files on the Snowball.
For example, when generating JPG files, the throughput started at 120 d/s for a few 100,000, but the performance progressively dropped and the throughput is now completly erratic between 35 and 50 docs/s.

The limitation is not the CPU:

• without output in memory I have 16,668 docs/s
• with local FS I have more than 120 docs/s

The idea is to leverage the snowball archive as described in this page:

• pack a few hundreds small files in a big archive file
• add a meta-data to tell S3 to uncompress the file at the storage level