agent: support scaling cooldowns to aid oscillation migration.

[jrasell]

Cooldown is now supported within a scaling policy as a time
duration. If the scaling policy omits the cooldown field, a
default value will be used. The default value can be controlled
by the CLI and has a default value of 5mins.

In the event the agent successfully registers a scaling action,
the policy manager will be instructed to place the policy eval
handler in cooldown by updating the ticker to the cooldown time
duration. The ticker is reset to its eval time duration on the
first tick out of cooldown.

Target plugins can optionally return a meta entry to detail the
last scaling event from their prespective. This allows the
autoscaler to enforce cooldowns, even when scaling actions have
been triggered manually by operators. The flow is similar to that
mentioned above, the only difference being that the ticker update
is based on the calculated remaning cooldown time from the last
event timestamp.

closes #12

[lgfa29]

A few comments, but nothing major. The most important one is the mutex for the ticker, I think we can have a race condition without it (but that was already true before this PR).

[lgfa29]

Much clear solution now. There's no state to keep track of anymore since the Go routine will block during cooldown, so it reduces complexity and mental loading when reasoning about the code.

I made a few suggestions but nothing major.

Good stuff 👍

[cgbaker]

some minor comments, no big changes.