Like #1429, but less wrong.

My prior PR misinterpreted certain paragraphs of Intel's processor documentation. Unaligned lock cmpxchg8b maintains atomicity regardless of alignment; this is not emulatable with ARM atomics alone. Properly emulating this requires a complicated locking setup that would almost certainly harm the performance of well-behaved x86 processes. In fact, I'm pretty sure the prior PR never worked - I've noticed this crash doesn't always occur when you launch mariadbd; and it's entirely possible that I was just getting very lucky with heap allocations that just so happened to align whatever packed structs are there.

So, instead, here's a new PR that triggers a segfault on unaligned atomic access. This will keep the user from losing any other work they might have been doing in other guest processes.