Sparse/Dense decision should be made per-row, not per-file

Marking this as beta 1 in the hope I have time to rustle it up and get it reviewed beforehand. If I do not, I will let it slide, but our behaviour right now is not brilliant for workloads with a variance in density, and it should not be challenging to make a more targeted decision.

We can also make use of CASSANDRA-9894 to make column encoding more efficient in many, even dense, cases.