ykd_opt_off test fails very occasionally

[ltratt]

Once in a while (I've now seen it twice, including on b13 when trying to merge #1585 -- it failed the first time), the innocuous-seeming ykd_opt_off test segfaults. It's not at all obvious to me why, and I'm not even sure how to recreate the crash. At least one case was in a BB build with RUST_TEST_SHUFFLE=1 cargo test. Is the shuffle aspect critical? I have no idea!

Having looked at the jit-asm output, the "normal" parts of the codegen are very short and simple, and look correct to me. I'm less sure about the sidetrace aspects or deopt.

So in a sense this is a terrible bug report: but there is something wrong that can be triggered by this tiny test, and sidetraces and deopt are my initial guesses as to what.