E2e shows occasional kubectl failures here, so add some retries.  We may want
to make this more general, but I think we should try it out in small scope
first.

Also clean up the retry loop so it doesn't process errors as successful runs
(discovered in testing).

Also simplify a bit of go template syntax.

Testing: I made kubectl randomly fail 50% of the time ($RANDOM%2 ==0) and
iterated until this gave me more helpful results.  Still not perfect, but
better.