perf: enable rocksdb prefix seek mode to improve hashkey scan.

[neverchanje]

Currently, our storage engine uses rocksdb's default options to construct a bloom filter, which only checks the existence of "hashkey+sortkey". For the situation where users scan over the range of hashkey (multigetrange, hashscan), the bloom filter makes no optimization if the hashkey does not exist. And rocksdb will search the corresponding SST data block, until it confirms the inexistence.

So what can we do with it? RocksDB provides an option called prefix seek mode. Its major improvement is adding prefix (the hashkey) as well as the whole key to the bloom filter. (https://github.com/facebook/rocksdb/wiki/RocksDB-Bloom-Filter#prefix-vs-whole-key)

For point lookups (multiget/get), only the whole key blooms are used. For range queries (multigetrange, hashscan, unorderedscan+hashkey prefix filtering), rocksdb responds without searching SST after the hashkey misses in the bloom filter.

One other improvement is to use hash-based memtable. Since Pegasus doesn't support cross-partition ordering, this implementation can be taken into consideration.