Parallel safety docs for CTEs

Hello,

A colleague noticed today that the docs still say that "Scans of
common table expressions (CTEs)" are "always parallel restricted".

While I think that strictly remains true at the implementation level,
from a user's perspective I think that's not been true since the
change to default to trying to inline CTEs rather than defaulting to
materializing them.

Attached is a patch to slightly modify the language; would be happy to
hear suggestions on a better way to improve this.

Regards,
James Coleman