CI: test each pull request once, trim Windows jobs; regenerate the single header for 5.0.2 - #2902

… jobs

- Workflows that ran on every push and every pull request now run on
  pushes to master and on pull requests only. A branch of this repository
  with an open pull request was tested twice. All of them can also be run
  by hand (workflow_dispatch).
- Add a concurrency group to every workflow: a new push to a pull request
  cancels its previous run. Pushes to master and scheduled runs are never
  cancelled.
- LoongArch64, Emscripten and Ubuntu 22.04 clang 13/14 now run on master
  only (or by hand).
- Windows: 30 jobs down to 19. vs18-clang-ci-cxx20.yml was an exact copy
  of vs18-clang-ci.yml (same name, no C++20 flag); it is now a single
  ClangCL C++20 build. RelWithDebInfo jobs are dropped, as are a few
  redundant shared/static and NaN/Inf combinations.

… comma_delimited

The json_sequence and comma_delimited stage-1 filters drop RS markers and
root-level commas by compacting structural_indexes in place. That left a stale
index in the slot past the compacted end, where stage 1 had stored the input
length, and the final batch copies that slot into the bookkeeping that
truncated_bytes() reads. Complete streams therefore reported spurious
truncated bytes: e.g. 2 for the comma-delimited input `1,2,3`, or 3 for the
json_sequence input `\x1e{"a":1}\n\x1e[2]\n`.

Restore the sentinel after compaction in both filters (shared by every kernel,
including fallback). Truncated tails are still reported: `{"a":1},{"b":`
yields 5. The extra store has no measurable cost (json_sequence and
comma_delimited parse_many throughput within noise on Apple M4 and on Xeon
Gold 6548N).

Update the iterate_many/parse_many docs and the document_stream headers: the
value is now meaningful in every format, under the same conditions as
before (iterate to the end, no document error). Add a regression test and
regenerate the single header.