Consider again propagating dynamic native dependencies

[alexcrichton]

In 2b9c774, we stopped propagating link arguments across crates. The main reason for this is that it was impossible to have a static library build dependency because the link arguments would always be propagated (and the library is only available during the build process).

Now that we have a proper linkage system set up (after #10528), we may want to consider re-propagating these dependencies. I have seen many failed compilations due to this not occurring (one is in issue #10543), and it seems reasonable to do this.

Because we would only do this for native dynamic dependencies, the dependency must already be in place to use the upstream crate, so why not just link it to the downstream crate?

I'm personally in favor of this (due to having a distinction between dynamic/static native libraries now), and I believe that this should be done.

Nominating for discussion.

[alexcrichton]

It appears that @luqmana has run into this when building rustdoc as well: https://gist.github.com/luqmana/a418261dc179db131770

[alexcrichton]

And again @luqmana found this building rustc: https://gist.github.com/luqmana/3eb91d9181cacf11cfbe