We should only set `deny(warnings)` in doctests when `deny-warnings = false`

[RalfJung]

deny(warnings) is usually not something we want to hard-code in the sources, it should only be set on CI -- otherwise lints can introduce frustrating roadblocks when they don't have to. I encountered this as part of this recent problem:

rust/library/core/src/lib.rs

Line 50 in 19e02c8

test(no_crate_inject, attr(deny(warnings))),

Would it be possible to set this flag inside ./x instead, and make it depend on the deny-warnings setting in bootstrap.toml? That would make it more consistent with the main crate builds as well. It seems this could be achieved by setting RUSTDOCFLAGS="--doctest-build-arg -D warnings -Zunstable-options".

[RalfJung]

Originally I thought we could use --cap-lints instead, but there is another test that checks that some code triggers a warning.

So after exploring quite a few options for how we could make things less fragile than they are right now, I conclude that getting rid of this hard-coded deny(warnings) is the only proper fix here, I think, to avoid spurious lints in a doctest from breaking miri-test-libstd.