Drop references to the former compiler from two comments

The Doxyfile had never run in CI, and its warnings went to a log file.

- It reads the headers only (*.hpp): Doxygen cannot match the explicit
  specializations of StringMaker and the PImpl constructors defined in the
  sources to their declarations, which gave ten of its twelve warnings.
- src/Progmasoft/Catch3/Internal/Preprocessor.hpp is left out: Doxygen reads
  priority_tag<N> deriving from priority_tag<N - 1> as a recursive class.
- WARN_AS_ERROR = FAIL_ON_WARNINGS, warnings on standard error.
- DOT_GRAPH_MAX_NODES = 200, so the include graph of All.hpp (103 nodes)
  is drawn; DOT_MULTI_TARGETS, which Doxygen 1.18 no longer knows, is gone.
- .github/workflows/doxygen.yml installs Doxygen and Graphviz, builds the
  reference and keeps the HTML.

Claude-Session: https://claude.ai/code/session_01Ji7xZMYjc4zankoc7hGBE1

Release Catch3 0.4.0

The documents keep Catch2's style, an anchor at the top of each page and
a contents list whose lines end in <br>; Codacy's markdownlint reported
the new contents line of the release notes as inline HTML.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ji7xZMYjc4zankoc7hGBE1

Every place that states the version says 0.4.0, written by
bazelisk run //tools:version -- 0.4.0, and the single-file distribution
carries it. The release notes describe the release: per-test coverage
with --coverage, secret-looking variables withheld from workers unless
--pass-env names them, the repository tools in C++17 under Bazel, and
versions with three numbers again.

Claude-Session: https://claude.ai/code/session_01Ji7xZMYjc4zankoc7hGBE1

Number versions MAJOR.MINOR.PATCH in the version tool

The macOS workflow and RepositoryReport.hpp explained themselves by another project; they now say what they do: macOS gets one job, and the JSON report uses the field names of RepositoryReport in camelCase.

Claude-Session: https://claude.ai/code/session_01Ji7xZMYjc4zankoc7hGBE1

NextPatch added one to the patch without a check, so a patch at
std::numeric_limits<unsigned>::max() wrapped to 0 and --next would have
moved the version back. It returns nothing there now; tests cover a
three- and a four-number version at the limit.

Claude-Session: https://claude.ai/code/session_01Ji7xZMYjc4zankoc7hGBE1