this was a pre-existing latent bug, we just didn't have any CI builders exercising it.

fixes the following errors:
```
 ---- [ui] tests/ui/sanitize/new-llvm-pass-manager-thin-lto.rs#opt1 stdout ----

error in revision `opt1`: test compilation failed although it shouldn't!
status: exit status: 1
command: "/checkout/obj/build/x86_64-unknown-linux-gnu/stage2/bin/rustc" "/checkout/tests/ui/sanitize/new-llvm-pass-manager-thin-lto.rs" "-Zthreads=1" "-Zsimulate-remapped-rust-src-base=/rustc/FAKE_PREFIX" "-Ztranslate-remapped-path-to-local-path=no" "--sysroot" "/checkout/obj/build/x86_64-unknown-linux-gnu/stage2" "--target=i686-unknown-linux-gnu" "--cfg" "opt1" "--error-format" "json" "--json" "future-incompat" "-Ccodegen-units=1" "-Zui-testing" "-Zdeduplicate-diagnostics=no" "-Cstrip=debuginfo" "--remap-path-prefix=/checkout/tests/ui=fake-test-src-base" "-o" "/checkout/obj/build/x86_64-unknown-linux-gnu/test/ui/sanitize/new-llvm-pass-manager-thin-lto.opt1/a" "-A" "unused" "-Crpath" "-Cdebuginfo=0" "-Lnative=/checkout/obj/build/i686-unknown-linux-gnu/native/rust-test-helpers" "-Clinker=x86_64-linux-gnu-gcc" "-L" "/checkout/obj/build/x86_64-unknown-linux-gnu/test/ui/sanitize/new-llvm-pass-manager-thin-lto.opt1/auxiliary" "-Zsanitizer=address" "-Clto=thin" "-Copt-level=1"
--- stderr -------------------------------
error: linking with `x86_64-linux-gnu-gcc` failed: exit status: 1
   = note: x86_64-linux-gnu-gcc: error: /checkout/obj/build/x86_64-unknown-linux-gnu/stage2/lib/rustlib/i686-unknown-linux-gnu/lib/librustc-nightly_rt.asan.a: No such file or directory

failures:
    [ui] tests/ui/sanitize/badfree.rs
    [ui] tests/ui/sanitize/address.rs
    [ui] tests/ui/sanitize/use-after-scope.rs
    [ui] tests/ui/sanitize/new-llvm-pass-manager-thin-lto.rs#opt0
    [ui] tests/ui/sanitize/new-llvm-pass-manager-thin-lto.rs#opt1
```

by using `rustc-builds-alt` if download-rustc is set

this also changes the download code to use a separate build/cache/ directory and .rustc-stamp stamp file depending on whether assertions are enabled.

…-error.rs

Not only are they a pain to update, but they differ depending on whether `parallel-rustc` is enabled or not

This was not the correct fix. The problem was two-fold:
- `download-rustc` didn't respect `llvm.assertions`
- `rust-dev` was missing a bump to `download-ci-llvm-stamp`

The first is fixed in this PR and the latter was fixed a while ago. Revert this change to avoid breaking `rpath = false`.