Make binary searches friendlier to mechanical harddisks

[alnkesq]

In order to store many years of data (1 year ≈ 1 TB, assuming no growth), some users might prefer using a cheaper mechanical harddisk.

The current implementation however uses many random accesses for binary searches.

We can create a small cache file for each ImmutableMultiDictionaryReader that can be fully loaded into RAM, that contains the start key of every (for example) 1024*1024 key block, that we use to perform an initial approximate binary search without needing to touch the harddisk.