EliasFano.Build: less iterations in upperBits loop

[AskAlexSharov]

Idea: instead of iterating over bits, iterate over words and get first-non-zero bit in word

Details: all Build() functions use naive O(64) bit scanning per word. Replace with bits.TrailingZeros64 + word &= word-1 to iterate only over set bits directly. This compiles to a single TZCNT instruction and eliminates all the skipped-zero-bit iterations.

Result: 2.75x speedup. Also speedup scales with sparsity of upperBits — at ~34% density, the inner loop runs ~3x fewer iterations

cpu: AMD EPYC 4344P 8-Core Processor
100 elements sequence: 233.0 ns/op -> 81.98 ns/op
1M elements sequence: 2344235 ns/op -> 824938 ns/op