…nge (-1, 1) to test (-4, 4). Cleanup code/comments. Test performance for all approximations.

* Partially apply clang-tidy fixes we don't use yet

- Put a bunch of stuff into anonymous namespaces
- Delete some redundant casts (e.g. casting an int to int)
- Add some const refs to avoid copies
- Remove meaningless inline qualifiers on in-class definitions and
constexpr functions
- Remove return-with-value from functions returning void
- Delete a little dead code
- Use std::min/max where appropriate
- Don't use a variable after std::forwarding it. It may have been moved
from.
- Use std::string::empty instead of comparing length to zero

* Undo unintentional formatting change

* Restore some necessary casts

* Add NOLINT to silence older clang-tidy

LLVM as it is built on the buildbots depends on `-lrt`, which is not a
target. Filter out non-target dependencies from consideration.

GCC 12 only supports _Float16 on x86. Support for ARM was added in
GCC 13. This causes a build failure in the manylinux_2_28 images.

The instructions for which llvm to acquire were stale

* Update pip package metadata

* Link to the CMake package docs from Doxygen

* Fix invalid Doxygen annotation in Serialization.h

PyPI rejected this because of a spacing issue.

A few quirks in the Markdown parser were worked around here.

The most notable is that the sequence `]:` causes Doxygen to interpret a
would-be link as a trailing reference even if it is not at the start of
a line. Duplicating the single bracket reference is a portable
workaround, i.e.

    [winget]  ~>  [winget][winget]

It also doesn't stop interpreting `@` directives inside inline code, so
it warns about our use of the `@` as a decorator symbol inside
Python.md.

Someone was using this as a reference expert schedule, but it was
stale and a bit simplistic for large matrices. I rescheduled it
to get a better fraction of peak. This also now demonstrates how
to use rfactor to block an sgemm over the k axis.