hammer2 - Refactor bulkfree

hammer2 - Refactor bulkfree

* Change the bulkfree scan from breadth-first to depth-first.  This
  improves disk performance significantly (~2x) and is also needed for the
  duplicate-detection feature.

* Create an 8-way set-associative hash table similar to what the live
  dedup code uses.  Record the data_off for elements we have processed
  and detect if a duplicate is encountered so we do not have to re-process
  the duplicate subtree.

  Also prioritize the table based on the aggregate bottom-up inode count
  to reduce the chance that a top-level duplicate (aka snapshot) will get
  kicked out of the hash table.

* Clean up the hammer2_chain_scan() API, making it more bref-centric
  which allows us to avoid instantiating chain structures for leaf
  entities.  This significantly improves performance and increases
  flexibility.

* Manual page adjustments for kern.maxvnodes settings suggestions.