Don't touch the lookaside cache if we're responding with a lookaside
cache entry. Also, leave the null entry behind if we're deliberately
dropping a request (a rare case) so that we don't have to process it
again. Fixes several lookaside problems in 1.10:
* When dropping a request because it was already being processed, we
were erroneously removing the null entry, causing us to process the
request again upon a second retransmit.
* When responding to a finished request with a lookaside entry, we
were removing and re-adding the entry to the cache, resetting its
time and performing unnecessary work.
* We were not caching responses we couldn't deliver because they were
too big for UDP, causing us to re-process the request when it came
in again via TCP instead of simply delivering the cached response.
Commit By: ghudson