There's no need to have rexi_server handle every process exit. There's
already an exception handler and workers can report on their own exit
reasons selectively. This also changes job tracking into a concurrent
ets update so if an RPC worker doesn't die in rexi code then rexi_server
will only handle half of the message passing load since we no longer
have to handle process exits.

This is fairly expensive for something that's not used very much. And
generally speaking when it could be useful things are too busy to really
rely on it as the messaging rates through rexi are so high.

This is a remnant of a prior iteration. Removing for simplification.

This changes cache eviction from a count of entries to the total amount
of memory used by the cache (as approximated by the memory used by ets
for the ?CACHE table). This behavior reverts to match the thresholds
used by the previous version as well as let users make more direct
choices on how large to make their design document caches.

Turns out we could go to evict an entry just after it had decided to
leave the cache (i.e., its ddoc was just deleted). When that happens
ddoc_cache_lru would die trying to shut the process down becuase the
process died normal earlier than expected.

Tests that modify an existing design document and then immediately try
and use it are a prime race condition against the cache invalidation.