memory defragmentation improvements

[romange]

1. Currently our defragmentation procedure is fully automatic but can sometimes miss important cases.
   We should add "MEMORY DEFRAGMENT" that will force kicking off the defragmentation process on all the data shards.
2. In addition, the defragmentation is triggered based on "used memory/max memoryt" threshold which is not accurate. Some workloads may have "used" part below the threshold and still have high RSS usage with high fragmentation rates for the data. We can not just call ReadShardMemUsage without preliminary cheaper checks. Therefore I suggest replacing the "used" check with RSS check, i.e when the RSS /maxmemory is high - to check the fragmentation stats. Finally, even in this case we may find ourselves calling ReadShardMemUsage too much. We should make sure we do not call it more often then once in K seconds.

[romange]

for (2) I am talking about mem_defrag_threshold flag

[romange]

for (1) to kick-off DoDefrag process manually without checks