v0.5.2

Version 0.5.2.

First, a correction. The last release went out tagged v5.01, which was
meant to be 0.5.1 and wasn't, and it left Booth looking four major
versions further along than it actually is. It isn't. This release puts
the numbering back where it belongs, and sorry to anyone who pinned the
old one or took the version at face value. The tag stays where it is so
nothing breaks underneath you, but the compiler now reports what it is.

The theme this cycle, without meaning to be, was the compiler telling
the truth. Semantic errors used to be printed and then ignored by every
mode except `--sema`, so the backend ran on source that had already been
rejected, wrote an output file and exited zero. Asking for several
backends at once wrote all of them over the same `-o` path and left you
whichever finished last, under the name you chose, again exiting zero.
Metal quietly narrowed a double-precision kernel to float and said
nothing, which is a real problem if you were counting on the precision.
`--amdgpu` ignored `-o` entirely and mixed a diagnostic into the
assembly on stdout. All four are fixed, and all four had been sitting
there being cheerfully wrong for a while.

The structural change is the backend contract. Every target now sits
behind a `be_desc_t` and registers itself in one list, and a backend
owns its own command line rather than reaching into a shared config
struct and the driver's argument loop. Adding a target used to mean
reading 420 KB of AMD backend to work out what was expected; now it
means reading one header and copying the skeleton. `main.c` lost about
a third of its length in the process, and the frontend stopped needing
to know what an AMD target enum is.

Booth also installs now. `make install` puts `kath`, the message
catalogues and a CMake package config into a prefix, so a downstream
project can `find_package(Booth)` and compile kernels as part of its own
build with `booth_add_kernel()`. There is a worked example under
`examples/cmake/` and CI builds it against a staged install on every
push, along with a check that the target list in the package config
hasn't drifted from what the backends actually accept.

Getting Booth no longer requires being able to build it. Every release
now carries prebuilt binaries for Linux, macOS and Windows, statically
linked so they have no runtime dependencies whatsoever, with checksums.
Unpack and run. This mattered more than I realised: the Windows build
had been quietly depending on MinGW's `libssp-0.dll`, so handing someone
the binary would not have worked on a machine without a toolchain, which
is exactly the machine they wanted it for.

There are coverage numbers for the first time, 74.1% of lines and 58.4%
of branches, reported by CI on every PR. That immediately turned up the
SSA register allocator having never been executed by a test at all, and
the six fixtures now pinning its behaviour also pin a real bug in it,
which is at least honest.

Thanks to Jorge Galvez, whose do-concurrent ocean benchmarks found three
genuine frontend bugs in an afternoon, and who let me test against his
code. Thanks to @maou3434 for the bare HIP warp and lane intrinsics,
which is their first contribution here and a very welcome one. And
thanks to @FileDelta for asking a simple question about the runtimes
that turned over considerably more than either of us expected.

The full itemised changelog for this release is in [CHANGELOG.md](https://github.com/Zaneham/Booth/blob/master/CHANGELOG.md).

Binaries below need no dependencies and no toolchain. Unpack and run. `SHA256SUMS` has the checksums.