constant: correctly handle signedness in const_to_opt_u128.

[eddyb]

One of the more insidious bugs found via Rustlantis (using a sample from @FractalFir's Rust-GPU port of it).

const_to_opt_u128 is used by rustc_codegen_ssa to optimize switchInt on the fly (i.e. replacing match CONST {...} with an unconditional branch to one of its cases), and that was working fine for unsigned discriminants, but we were wrong for signed ones, due to MIR storing switchInt cases as zero-extended-to-u128 (e.g. -1i8 => ... becomes 255 => ...).

More specifically, when const_to_opt_u128(x, /*sign_ext*/ false) was called with x of a signed integer type, we were wrongly sign-extending (based on type) instead of zero-extending (based on sign_ext == false).

---

TODO: still needs a test, perhaps something along the lines of:

#[inline(never)]
fn const_i32_is_neg1<const X: i32>() -> bool {
    match X {
        -1 => true,
        _ => false,
    }
}

fn main() {
    assert!(const_i32_is_neg1::<-1>());
}

(you can see the lack of a switch in the LLVM IR: https://godbolt.org/z/hj6WEKj3v)
_{(also, unrelatedly, that godbolt sample made me notice 1.94 -> 1.95-beta generates more unnecessary LLVM IR in some situations, and nightly still does, too)}