Avoid unaligned writes in CPU RandomBits

[markjouh]

Summary

• avoid unaligned 32-bit stores in the CPU RandomBits implementation
• preserve the generated bytes and Threefry sequence by using memcpy for complete words
• leave the existing partial-word copy path unchanged

Root cause

For batched keys, each key's output begins bytes_per_key bytes after the
previous one. When that stride is not divisible by four, casting the start of a
later key's output to uint32_t* and writing through it is undefined behavior.

The existing test random bits case covers an affected layout: vmapping five
uint16 values over two keys makes each per-key output 10 bytes wide. Before
this change, that test aborts under UBSan with a misaligned uint32_t store.

Validation

• UBSAN_OPTIONS=halt_on_error=1:print_stacktrace=1 build-ubsan/tests/tests -tc='test random bits' (1 test, 28 assertions)
• release C++ suite (244 tests, 3,238 assertions)
• UBSan suite excluding two unrelated pre-existing GGUF alignment failures (242 tests, 3,187 assertions)
• clang-format --dry-run --Werror mlx/backend/cpu/primitives.cpp
• git diff --check