[lemire]

Fixes #80.

binary_fuse8_populate cast i * size to 32 bits before shifting by blockBits, so the product wrapped once it reached 2^32 (from about 8.4M keys). Wrapped block starts land inside the low blocks, so placement walks many blocks before finding a free slot. The filter stayed correct, but populate got much slower. This applies the parenthesization already used in binary_fuse16_populate:

-      startPos[i] = (uint32_t)((uint64_t)i * size) >> blockBits;
+      startPos[i] = (uint32_t)(((uint64_t)i * size) >> blockBits);

At 20M keys (-O3, Apple silicon, one run), fuse8 populate goes from 120.9 ns/key to 27.2 ns/key, and every key is still contained. make unit && ./unit passes.