Linking modifier `whole-archive` and link time `cfg` are not supported on bundled static libraries

[petrochenkov]

This is more or less a part of #81490.

If you try to combine +whole-archive with (possibly implicit) +bundle you will encounter this error: "linking modifiers +bundle and +whole-archive are not compatible with each other when generating rlibs".
The whole-archive modifier is significantly limited due to this issue and cannot be used in foo-sys rlibs bundling their native libfoo.a dependencies, which is one of the main use cases.

We also have to enable the +whole-archive modifier implicitly in one ad hoc case (https://github.com/rust-lang/rust/blob/master/compiler/rustc_codegen_ssa/src/back/link.rs#L2209-L2213) for backward compatibility, because it cannot be enabled manually.

cfg for bundled static libraries was implemented just enough to make the case of musl libc bundled into libc crate to work, but it doesn't work in general case.
(Note that musl libc no longer uses bundling, see #72274.)

Also see #88085 and #88161 for some discussion related to this issue.

[petrochenkov]

We need to make bundled libraries behave more like regular static libraries libfoo.a.

The simplest way, which is also most compatible with traditional tooling is to literally move them out of rlibs to separate .a files, like this

.
├── libfoo-sys.rlib
└── libfoo-sys.rlib.bundle
    └── libfoo.a

It solves all the issues above:

• libfoo.a can be passed to linker as a separate static library, with preserving all the static library semantics, thus it can be
  • wrapped into -whole-archive ... -no-whole-archive
  • not passed at all if some relevant link time cfg is false
  • moved to some other place on the command line if necessary, not necessarily next to libfoo-sys.rlib
• it's also fast, rustc doesn't have to extract object files to some temporary .a file on the fly or something like that, when static library semantics are necessary.

The problem is that we'll have to teach cargo to track these new extra files.
Third party build systems will also need to support this (although it's possible that they never use bundling and always link their native libraries in their own way).

If the separate file solution doesn't work out we'll need to figure out how to bundle object files into rlibs in some smarter way, with preserving their old structure, and then how to repack them on the fly if they need to be represented as a static library.