Improve GC scalability on TarMK

This issue is about making TarMK gc more scalable:

• how to deal with huge repositories.
• how to deal with massive concurrent writes.
• how can we improve monitoring to determine gc health.
  • Monitor deduplication caches (e.g. deduplication of checkpoints)

Possible avenues to explore:

• Can we partition gc? (e.g. along sub-trees, along volatile vs. static content)
• Can we pause and resume gc? (e.g. to give precedence to concurrent writes)
• Can we make gc a real background process not contending with foreground operations?

This issue is a follow up to OAK-2849, which was about efficacy of gc.