[ML] bucket_count is inaccurate when there are gaps in the data

[elasticmachine]

Original comment by @davidkyle:

Open a job send some data and close the job then reopen the job and send some data timestamped a week later than the previous batch. Autodetect will create empty bucket results for the intervening period but DataCounts::bucket_count will not reflect that.

The testMlBasicMultiNodeIT::testMiniFarequoteReopen does exactly this but the test was asserting that bucket_count == 2 rather than bucket_count = 7 days of buckets. bucket_count should equal to the number of buckets written by autodetect, with the caveat that old results are sometimes pruned.