Bugfix: maybe_placeholders pads shorter alternatives of an optional (#1078)

[gaoflow]

maybe_placeholders promises an optional [...] group a fixed number of children so consumers can index positionally. When the group's alternatives differ in width, the arity varied per input:

!start: "a" ["b" | "c" "d"]
"a"   -> ['a', None, None]
"ab"  -> ['a', 'b']          # arity 2 -> children[2] raises IndexError
"acd" -> ['a', 'c', 'd']

maybe() padded only the absent case up to the widest branch; a matched narrower branch was left short. This is the case flagged # Not implemented; current behavior is incorrect in test_maybe_placeholders, and issue #1078 ("the argument count must remain the same whether or not there is a match").

The fix pads each shorter alternative with trailing placeholders during EBNF expansion, so every derivation of the optional has the same arity — "ab" now yields ['a', 'b', None]. It also handles nested and non-leading alternations ([("a" | "b" "c") "d"], ["a" | "b" "c"] "z"), and leaves maybe_placeholders=False untouched. Same result on earley and lalr.

Re-enables the maintainer's own commented expectation and adds the sibling cases as regression tests.