cg_LLVM: Stop needing an alloca for volatile loads This ended up also being reimplementing it to not use `load` of the `llvm_type`, since without doing that everything blew up horribly. cc the zulip conversation [#t-opsem > Defining volatile splitting @ 💬](https://rust-lang.zulipchat.com/#narrow/channel/136281-t-opsem/topic/Defining.20volatile.20splitting/near/597451615). And while I'm here, improve the tests to check that the unaligned ones are actually unaligned, since `unaligned_volatile_load::<u8>` doesn't actually test anything. r? @nikic cc @RalfJung MCP tracking issue rust-lang/rust#153250