Row granularity data keys

Currently we can only set data keys at column family granularity. Consider support for setting data keys per row. Then tenant specific data keys in a multitenant table becomes possible.

HBASE-13512 proposes to do this per region, which is much easier to achieve than what is proposed here, since we'd only vary the existing policy for assigning data keys to whole HFiles. However it seems not that difficult to do this per row.

When encrypting HFiles we apply encryption on a per block basis, so per row encryption can be achievable if internal to an HFile we are willing to pack cells into blocks by row. With per-row encryption active we would start a new block after we're done with cells for a given row. The result would be an HFile with blocks of varying size each including a data key reference in the block header, and presumably the set of relevant data keys stored in a meta block. Looks like the HFile writer code will let us do that readily, but performance concerns and knock-on effects in block caching could make things interesting. Another complication is indexing. If we can say that keys themselves are not sensitive, then we don't have to do anything special about block indexes. If we cannot say that keys themselves are not sensitive, offhand I don't know how to do block indexes, sure we could encrypt leaves of the index corresponding to specific data blocks independently, but what to do higher in the tree?