PackedInts.fastestFormatAndBits returns slower format for Direct [LUCENE-5352]

from some simple testing, DirectPacked64SingleBlockReader seems slower than DirectPackedReader (at least in a few cases i tested, more testing needed, but look at the code).

this is unintuitive, because by passing e.g. FASTER you are making it slower.

---
Migrated from [LUCENE-5352](https://issues.apache.org/jira/browse/LUCENE-5352) by Robert Muir (@rmuir), 1 vote

Start at PackedInts.fastestFormatAndBits and compare DirectPacked64SingleBlockReader with DirectPackedReader, focusing on the FASTER selection described in the issue. Reproduce the simple performance cases first, then determine whether the format choice is incorrect; done means the reported slower selection is resolved or its behavior is supported by measured results.

Written by the indexing model from the issue text.