Match PETSc object leaks on the syntax tree, not on source text

[filoferra]

audit_petsc_file paired object creations with destroy calls by running seven near-identical regexes over each source line. That approach had four defects, all of which the fixture in test/fixtures/leaky.jl now covers:

• each regex block ended in continue, so only one event per line was recorded; v = VecCreateSeq(...); VecDestroy(petsclib, v) counted the creation and missed the release
• only whole-line comments were skipped, so # destroy!(dm) and a destroy mentioned inside a string literal both counted as releases
• only bare and PETSc.-qualified destroy were recognised, so LibPETSc.VecDestroy read as a leak
• DMPlex objects were not tracked at all, and the high-level Mat/Vec factories (MatSeqAIJ, MatSeqDense, VecSeq, ...) were missed because their names carry no Create or Duplicate marker

Walking the parsed AST collapses the seven blocks into one traversal over two name sets. Destroy calls now record every plain name in the argument list, since the object is the first argument high-level (destroy!(v)) and the second low-level (VecDestroy(petsclib, v)).

Note: audit_petsc_file returns a NamedTuple of created, destroyed, finalized and leaked rather than nothing, so the behaviour can be asserted. It still prints the same report; pass verbose = false to suppress it. Nothing in the repo consumed the old return value.

[boriskaus]

@filoferra can you check the conflicts?