Add dictionary filter statistics

[sadikovi]

This PR adds dictionary filter statistics that are based on java.util.HashSet. Now user can choose between bloom (Bloom filter statistics) and dict (Dictionary filter statistics) when creating index.

Also added/updated filter statistics tests and read correctness tests.

Benchmarks on test dataset of 1,000,000 records with simple Or(EqualTo(code,2),EqualTo(code,120)) query shows that with bloom filter statistics Spark would scan ~8 files. Same dataset with same filter and dictionary filter statistics would read only 2 files (both codes are in different files).

Dictionary filters are slower (x2) to create than bloom filters, and they take more space, stats below for sample dataset:

bloom filters:
total: 6.6M
metadata: 300K
filter file: 2.2K

dict filters:
total: 16M
metadata: 300K
filter file: 9.9K - 19K

[coveralls]

Coverage increased (+0.1%) to 95.341% when pulling 003aafd on dictionary-filter-2 into 32f4a6c on master.