Resolves #438. There were two remaining uses of error_syntax:

• One is in the region parsing non-capturing groups and assertions. If it is not one of these, then ? is essentially a quantifier applied to nothing, so error_badrepeat seems an apt choice among the available codes to me. (The error code error_paren used by libstdc++ feels a bit weird to me in this context because this error code is intended for unbalanced parentheses, which this isn't.)
• The other case can only be reached by an STL or compiler bug, so I replaced it by _STL_INTERNAL_CHECK: _Alternative() only returns normally when _Mchar is _Meta_eos, _Meta_bar or it is _Meta_rpar while _Disj_count != 0. Its caller, _Disjunction(), only returns if _Mchar is not _Meta_bar after _Alternative() returned. So this means the error could only be thrown when _Mchar is _Meta_eos or _Meta_rpar with _Disj_count != 0. But in either case, this is a bug in the parser: Either _Meta_eos was assigned to _Mchar before the end of the regex was reached (but see _Trans() for the only case where _Meta_eos is assigned) or an unbalanced number of increments and decrements to Disj_count were performed.