[rr-]

Three changes to items in Google params sections (Args:, Attributes:, Keyword Args: and the like), and one to raises sections:

• A param's name alone on its line, x or x (int), with its description indented below it or none, is now an item. It used to be a missing-colon problem.
• A line at the items' indent that has no head, such as wrapped prose that wasn't indented, now goes on with the item before it instead of starting a bogus item or raising. This applies to raises sections too. A line counts as a head if it is names (no spaces, or a x, y list) with an optional (type), before a colon or alone on the line, or at most two words before a colon, as in total_attempts Optional[int]:, which google-auth writes. Only the first item of a section can still be a missing-colon problem.
• A colon inside the brackets, as in x (int, default: 3): text, no longer ends the head, so that param gets type int and default 3.

To keep round trips, compose now writes a description whose first line is indented further than the rest on the line below its head.

Case diffs:

• corpus/google-auth/..._get_cert_config_path: used to raise on a wrapped line, and now parses, with the wrapped line in the description. Auto now picks Google for it instead of ReST.
• corpus/rich/rich.cells._split_text: used to raise on cell_position Offset in cells., which has no colon. That line now continues text's description.
• google/google_typed_parameter_delimiters_05: value (int) followed by a line ): description is now a bare item with description ): description, instead of a head wrapped over three lines. This one is open to question, though the text is all still there.
• google/long_parse_error_01: the input was a single long word, which is now a bare name. I changed it to long words with spaces, so it still tests the error excerpt.
• test_try_parse: the two Google problem tests now put the line with no colon first, since a later one continues the item before it.
• test_invariants: generated 166 and 378 (Google, BODY) round trip now and leave the exclusion list.

Parity against origin/main: 854 of 29183 docstrings differ, case 3, stdlib 17, generated 834. 683 parse only in this tree (case 2, stdlib 9, generated 672), mostly bare names that used to be problems. None parse only in base, so no stdlib docstring newly makes strict parse raise. The stdlib diffs are mostly pydoc-style name -- text lists under Arguments:, which are now one item (or one problem) instead of one per line. Auto parsing changes for 589 docstrings, nearly all generated, because Google now parses them.