changefeedccl: add prototype of EXPERIMENTAL_CHANGEFEED

This implementation is a prototype that will be used to get early
feedback from users about whether the way we structure kafka
topics/partitions/encodings work for them.

It works by polling at a configurable interval (default 1 second). Each
poll picks a new highwater timestamp and sends a kv ExportRequest per
covered span. These requests are configured to return (inline in the
response) any kvs that were written between the previous highwater mark
and the new one. These responses are fed to a RowFetcher for conversion
to sql rows.

There is a 1:1 mapping between tables and which kafka topic they're
emitted to. The topic name is some prefix (an option in the sql command)
plus the table name. Each row is emitted as an kafka entry with key
unset and a json value mapping table field names to corresponding
datums. This format will almost certainly be tweaked in followup PRs but
this is sufficient to get everything working end-to-end with kafka
feeding into elasticsearch. Changed rows are emitted "at least once" and
may be duplicated if there are failures or restarts.

Changefeeds are kept as jobs and can be paused, resumed, and cancelled.

Current limitations:
- All changes are fed through a single node, limiting scalability and
  creating a hotspot on that node
- DELETEs don't work
- Schema changes are ignored
- Interleaved tables create duplicate entries
- Job progress is always 0% (bad ux)

Release note: None

pkg/ccl/ccl_init.go

@@ -15,6 +15,7 @@ import (
 	// ccl init hooks
 	_ "github.com/cockroachdb/cockroach/pkg/ccl/backupccl"
 	_ "github.com/cockroachdb/cockroach/pkg/ccl/buildccl"
+	_ "github.com/cockroachdb/cockroach/pkg/ccl/changefeedccl"
 	_ "github.com/cockroachdb/cockroach/pkg/ccl/cliccl"
 	_ "github.com/cockroachdb/cockroach/pkg/ccl/importccl"
 	_ "github.com/cockroachdb/cockroach/pkg/ccl/partitionccl"