Shift before truncating block starts in binary_fuse8_populate (fixes #80)

The cast to uint32_t was applied to i * size before the shift, so the
product wrapped above 2^32 (from about 8.4M keys) and block starts landed
in low positions, slowing populate several fold. Match the fuse16 code.

Author: lemire

include/binaryfusefilter.h

@@ -332,7 +332,7 @@ static inline bool binary_fuse8_populate(uint64_t *keys, uint32_t size,
     for (uint32_t i = 0; i < block; i++) {
       // important : i * size would overflow as a 32-bit number in some
       // cases.
-      startPos[i] = (uint32_t)((uint64_t)i * size) >> blockBits;
+      startPos[i] = (uint32_t)(((uint64_t)i * size) >> blockBits);
     }
 
     uint64_t maskblock = block - 1;