Rollback to a point which captures full snapshot in checkpoint mgr (but not persisted)

Currently we always rollback to a point that is persisted on the disk.

(http://src.couchbase.org/source/xref/trunk/ep-engine/src/kv_bucket.cc#2492,

http://src.couchbase.org/source/xref/trunk/ep-engine/src/vbucket.cc#1821)

If the rollback seqno requested is at a point on checkpoint mgr with unpersisted snapshots, then rolling back to a point which captures the last full snapshot is better (has lesser data loss) when compared to rolling back to a point that is persisted.

Example:

(numbers below are sequence numbers)

1 to 50 on disk.

51 to 60 to chkpt 1; 61 to 70 on chkpt2.

Say there is a rollback request to 64. We currently rollback to 50. (last snapshot persisted on disk).

This is not incorrect, but not as good as rolling back to 60 (last full snapshot).