hammer - Add simple I/O optimization, collect stats on new deduplications
* Add a simple I/O optimization. When dealing with CRC collisions scan
the SHA sub-tree looking for the data_offset/data_len match, allowing
us to avoid unnecessary I/O reads.
This effects data which has already been de-duplicated, so it will improve
dedup rescan times on the portion that has already been dedupped.
* Collect statistics on NEW deduplications during the pass and print them
out in the final stats as well as in the ^T stats.