This allows for a more optimal std.crypto.tlcsprng codepath.

Without it a an "incorrect alignment" panic is triggered from
crypto.tlcsprng which aligns a threadlocal but it's actually
not aligned, thus detected by the safety check.

It appears that LLVM-IR does attribute the storage with alignment
but it is ultimately not respected in the final binary for netbsd
and dragonfly.