Add table parsing fixup to merge row types

"rowspan" is unable to span across different "thead", and "tbody"
elements which make it not useful with the standard row parsing
routines. This commit adds a final step when parsing a table to join all
"thead" and "tbody" rows into just two set of elements.

This still has the limitation that a rowspan from a header won't
span to body elements, but this is not the most usual situation.

Signed-off-by: Fabio Utzig <[email protected]>

breathe/renderer/sphinxrenderer.py

@@ -1656,7 +1656,26 @@ def visit_doctable(self, node) -> List[Node]:
             colspec.attributes['colwidth'] = 'auto'
             tgroup += colspec
         table += tgroup
-        tgroup += self.render_iterable(node.row)
+        rows = self.render_iterable(node.row)
+
+        # this code depends on visit_docrow(), and expects the same elements used to
+        # "envelop" rows there, namely thead and tbody (eg it will need to be updated
+        # if Doxygen one day adds support for tfoot)
+
+        tags = {row.starttag(): [] for row in rows}  # type: Dict[str, List]
+        for row in rows:
+            tags[row.starttag()].append(row.next_node())
+
+        def merge_row_types(root, elem, elems):
+            for node in elems:
+                elem += node
+            root += elem
+
+        for klass in [nodes.thead, nodes.tbody]:
+            obj = klass()
+            if obj.starttag() in tags:
+                merge_row_types(tgroup, obj, tags[obj.starttag()])
+
         return [table]
 
     def visit_mixedcontainer(self, node) -> List[Node]: