Fix some inefficiencies in SQLite implementation

[gortiz]

This PR includes some improvements on SQLite implementation. I guess they could also be applied to H2, but being heap based, I don't think we care that much about H2.

With this changes results for sqlite went from 20 times slower than rocksdb to 30% faster:

Benchmark                                 (_keyCardinality)  (_metadataStoreType)  (_numRows)  (_numRuns)  (_numSegments)  Mode  Cnt    Score   Error  Units
BenchmarkUpsertAddRecord.upsertAddRecord            5000000               ROCKSDB        1000          10              10  avgt    2  237.424          ms/op
BenchmarkUpsertAddRecord.upsertAddRecord            5000000                SQLITE        1000          10              10  avgt    2  179.725          ms/op