Compiler bootstrap infrastructure and tests should be exercising -g more aggressively

[pnkfelix]

On #93006 a change landed that broke rustc -g for unsized tuples, in a manner that ICE'd the compiler.

We didn't notice that it happened until nearly two weeks later, when #93866 and #93871 were filed.

The reason I'm filing a bug is that we have tests of unsized tuples; even the libcore test suite itself uses them. (I.e., this bug broke my attempts to bootstrap Rust itself with debuginfo-level = 2 in the config.toml. But the test suite itself is not exercising -g enough to observe that combination.

At the very least, PR's that are making changes to debuginfo should be going through the process of bootstrapping with debuginfo-level = 2, as a sanity check.