Fix unaligned packed binary code access - #88

Summary

• write packed binary words with byte arithmetic and memcpy
• load packed words safely in AVX2, AVX-512, and HNSW mask kernels
• preserve the existing HNSW memory and serialized index layout
• add explicit two-word unaligned read/write regression tests

Verification

• full C++ suite: 48/48 passed
• focused UBSan tests passed for pack, AVX2/AVX-512 mask reads, and HNSW search
• formatting and git diff checks passed