fix: handle error recovery better for dedents in case a string start is possible

[amaanq]

I noticed this slight regression when running the tests in tree-sitter core - @maxbrunsfeld

This won't handle error recovery well for f-strings I think - which is what the inclusion of not dedenting in case string_start was valid was for anyways (they can be dedented and not count as a dedent)..

[maxbrunsfeld]

Thanks!