Since [1], the issue should be resolved so the workaround can be
dropped.

This reverts commit 88e83b9.

[1]: #759

Replace `public_test_dep!` by placing optionally public items into new
modules, then controlling what is exported with the `public-test-deps`
feature.

This is nicer for automatic formatting and diagnostics.

This is a reland of 2e2a925 ("Eliminate the use of
`public_test_dep!`"), which was reverted in 47e50fd ('Revert "Eliminate
the use of..."') due to a bug exposed at [1], reapplied in d4abaf4
because the issue should have been fixed in [2], then reverted again in
f6eef07 because [2] did not actually fix the issue.

[3] has landed in rust-lang/rust since then, which should resolve the
last problem remaining after [2]. So, apply this change for what is
hopefully the final time.

[1]: rust-lang/rust#128691
[2]: rust-lang/rust#135278
[3]: rust-lang/rust#135501

It seems as if `f16` works on MIPS now according to my
testing on Rust master with LLVM 20, and I was asked to
create PRs with my changes.

I only tested on the flavour of `mipsel-unknown-linux-gnu`
hardware that happens to be available to me, so I can't say
anything about other MIPS hardware, but from a casual
skimming of the LLVM code ([1], [2]) it seems like `f16`
should work on all MIPS hardware. So enable it for all MIPS
hardware.

[1]: https://github.com/rust-lang/llvm-project/blob/rustc/20.1-2025-02-13/llvm/lib/Target/Mips/MipsISelLowering.h#L370
[2]: https://github.com/rust-lang/llvm-project/blob/rustc/20.1-2025-02-13/llvm/lib/CodeGen/TargetLoweringBase.cpp#L1367-L1388

This requires privately reexporting `libm`'s `support` module at crate
root, where it is expected for macros. Once `libm` is made always
available, the reexport can be simplified.

This delta adds a lot of routines to `f16` and `f128`:

* ceil
* floor
* fma (f128 only)
* fmax
* fmin
* fmod
* ldexp
* rint
* round
* scalbn
* sqrt

Additionally, the following new API was added for all four float types:

* fmaximum
* fmaximum_num
* fminimum
* fminimum_num
* roundeven

There are also some significant performance improvements for `sqrt` and
`sqrtf`, as well as precision improvements for `cbrt` (both `f32` and
`f64` versions of this function are now always correctly rounded).