Do not store the lower bits mask in EliasFanoReader
Summary:
Computing the mask on access has negligible cost as it can be
hoisted out of the linear search loop, and furthermore on Haswell we
can use the the `BZHI` instruction.
I also experimented with `BEXTR` but it ended up being slower because
computing the pattern operand requires a shift and an or (it's
probably meant for when the pattern is precomputed).
Reviewed By: philippv
Differential Revision:
D4976657
fbshipit-source-id:
e4c4ca5f0a785595587e6d6ad4676f5b216291cf