If, for whatever reason, the request to fetch features data fails,
simply retry a little later (with a randomised exponential backoff).
This commit uses the [retry module](https://www.npmjs.com/package/retry)
to provide the backoff logic.

N.B. If we fail 10 times (over a period spanning between ~15 and ~30
minutes, using the defaults from the retry module) then we will simply
stop looking for new features data. There's no point burning client CPU
forever in the hope that the server reappears.

Fixes #2547.