A variant of SQL that can adapt to variable column rows should certainly be possible, much as SQLite implements a sql variant with variably typed columns.
Cassandra's CQL is an SQL variant/subset (albeit without joins, which don't scale out horizontally anyway) that is adapted to variable partitions with rows grouped into partitions. It addresses the problem of locality by embedding partition awareness into the query language.
Yes, that is the best example I've seen. But, by the same token (pun intended this time), it is only SQL-ish. It isn't SQL, which of course they are up-front about by calling it something else. Careful study and some practice usage of CQL can provide a practical object lesson in why "give me SQL!" is not always a reasonable demand to make of a datastore unconstrained by preconceived notions about semantics. SQL presupposes/affords more than meets the eye.