Make filesort support sorting a join of multiple tables, not just a single
table. The biggest change is that the sort types that depend on row IDs
need to support multiple ones instead of just one -- this means that we
can do such sorts without having to materialize them (StreamingIterator
already took care of most of the common cases, although with a fair bit
of unneccessary copying).

Filesort now takes in the estimate for number of rows (only used for
excluding priority queue, if there are so few rows that we assume multiple
buffers are not needed) from the estimates in the access path instead of asking
the table, since there's not necessarily “one table” anymore. For some tests,
this actually gives a more precise estimate, although it's likely to be fairly
inconsequential in practice.

Nothing in the executor currently sets up access paths with such multi-table
sorts, since fields are always resolved to the temporary table (so it's not
just a matter of changing), but the new optimizer will.

Change-Id: I5304035b1fa36b622e7d58d5464d6b3ec0e50ba8