Fix: Avoid unaligned serial memory accesses

[emecii]

Fix: Avoid unaligned serial memory accesses

Summary

• replace the serial copy and move paths' unaligned sz_u64_t* dereferences with an alignment-safe eight-byte transfer helper
• retain load-before-store ordering so overlapping sz_move_serial calls work in both directions
• exercise misaligned copy plus forward and backward overlapping moves in the existing C++ memory unit test

Fixes #279.

Testing

• SZ_TESTS_FILTER='test_memory_(unit|all|safety)' SZ_TESTS_MULTIPLIER=0.1 SZ_TESTS_SEED=279 /tmp/stringzilla-279-cpp20-ubsan
• SZ_TESTS_MULTIPLIER=0.1 SZ_TESTS_SEED=279 /tmp/stringzilla-279-cpp20
• direct Clang C++11/14/17 compilation of test/string.cpp

[ashvardanian]

Please check the surrounding lines next to your PR. There is already a built-in setting to avoid misaligned loads.