Close out #33: reduce_one() sanitizer crash, UBSan function-type reports, R-hub workflow

[mattfidler]

Closes #33.

Is #33 fixed?

Yes, by #32 (the first reduction path moved off a process-wide static onto reduce_one()'s stack). Verified locally with clang ASan+UBSan:

	1.3.1-13 (CRAN)	1.3.1-14 (dev)
concurrent dparse() test, 16 threads	parse.c:1397:57 … 0xbebebebebebebebe then SEGV (the issue's signature)	clean
rxode2 5.1.8 parallel CSE (.rxOptExprC, 1–8 threads)	same crash	clean

Only clang builds ever used the shared vector. In vec_add(paths, new_VecZNode(paths, ...)) the n++ and the call were unsequenced, and gcc took a fresh malloc()ed path. That is why only R-hub's clang containers crashed.

Changes

• src/parse.c: build_paths() calls new_VecZNode() before vec_add(), so every compiler takes the same path. With only this change applied to the old static-path1 code, the existing concurrent test now crashes it under gcc too.
• src/lex.c, src/write_tables.c: the hash/cmp callbacks take void * to match hash_fn_t/cmp_fn_t exactly. Clang's -fsanitize=function reported "call to function through pointer to incorrect function type" from mkdparse(), which would fail R-hub's UBSan step.
• .github/workflows/rhub.yaml (new, based on the r-hub v1 template):
  • sets NOT_CRAN=true so the concurrent test runs (it skips on CRAN);
  • after the check, installs this checkout's dparser plus nlmixr2/rxode2 and runs .github/rhub/rxode2-parallel-cse.R, failing on any sanitizer report;
  • the extra step can be turned off with the downstream input.
• NEWS.md entries.

Testing

• Full test suite under gcc ASan/UBSan and clang ASan/UBSan: 690 expectations, 0 failures, no sanitizer reports.
• The workflow can only be dispatched once it is on main; after merging, run rhub::rhub_check(platforms = c("clang-asan", "clang-ubsan")).