Every hook execution check_refresh_available is called, starting snapd

[Xav Paice]

The bootstrap of the snap layer calls `hookenv.atstart(check_refresh_available)` at the start of every hook execution, including update-status. This runs `snap refresh --list` (even if there's no snaps installed), and that causes at the least some noise in syslog, and a start of snapd plus whatever the refresh actually checks for.

This is likely not necessary for every hook execution - maybe better on config-changed only. It's expensive when there's a lot of LXD containers on a machine all running this same process every 5 mins. The operation looks to be taking approx. 5s each on my small test environment, a busy site may be more.

[Stuart Bishop (stub)]

We need to use hookenv.atstart() to get flags set before the main reactive loop kicks off. However, I see no reason the refresh has to run every update-status hook; it is setting the snap.refresh-available.* flags, but they can just as easily be several hours out of date as they can be 5 minutes out of date. I suspect these flags are just a bad idea and not being used.

We do want to do an unconditional refresh in the upgrade-charm hook, in case channels etc. defined in layer.yaml have changed.

I'm thinking the way forward is a layer.yaml option, enable_refresh_flags, defaulting to False, which simply causes check_refresh_available to do nothing. While backwards incompatible, it is a simple enough tweak by the charm author to turn back on if they are actually using the feature.