cc: #556

Channel.last_notify is the datetime of the most recent
notification sent via the channel. Channel.last_error is
the error message (blank if the delivery was successful).

In the integrations list, "Last Notification" column,
use these fields instead of looking up the most recent
Notification object. This saves some db queries,
and also fixes a subtle issue: if prunenotifications
cleans up all notifications for a given channel, the
"Last Notification" column would display "Never", which
would not be correct – not any more.

Profile.next_nag_date tracks when the next hourly/daily reminder
should be sent. Normally, sendalerts sets this field when
a check goes down, and sendreports clears it out whenever
it is about to send a reminder but realizes all checks are up.

The problem: sendalerts can set next_nag_date to a non-null
value, but it does not clear it out when all checks are up.
This can result in a hourly/daily reminder being sent out
at the wrong time. Specific example, assuming hourly reminders:

13:00: Check A goes down. next_nag_date gets set to 14:00.
13:05: Check A goes up. next_nag_date remains set to 14:00.
13:55: Check B goes down. next_nag_date remains set to 14:00.
14:00: Healthchecks sends a hourly reminder, just 5 minutes
       after Check B going down. It should have sent the reminder
       at 13:55 + 1 hour = 14:55

The fix: sendalerts can now both set and clear the next_nag_date
field. The main changes are in Project.update_next_nag_dates()
and in Profile.update_next_nag_date(). With the fix:

13:00: Check A goes down. next_nag_date gets set to 14:00.
13:05: Check A goes up. next_nag_date gets set to null.
13:55: Check B goes down. next_nag_date gets set to 14:55.
14:55: Healthchecks sends a hourly reminder.

To send notifications, sendalerts calls Flip.send_alerts().
I updated Flip.send_alerts() to be a generator, and to yield
a (channel, error, send_time_in_seconds) triple per sent
notification.

This avoids duplicate calls to Check.get_grace_start() in several places.