Using `driver_output_term()` in an async worker pool thread
is a big no-no.  A solution is to use `driver_output_term()`
in the `ready_async()` function, which executes in the proper
thread context.

Tested via:

* Run with a debug SMP build of the Erlang VM.  There are no
  `beam/global.h:1345: Lock check assertion` errors.
* Use the following to run a large number of Map/Reduce jobs to
  try to provoke a memory leak.

The following can be cut-and-paste'd into a Riak console/shell.
On my MBP, it runs for (very roughly) an hour.  Perhaps it should
run even longer to try to find very small memory leaks?

    {ok, C} = riak:local_client().
    [C:put(riak_object:new(<<"test">>, list_to_binary(integer_to_list(X)), list_to_binary(integer_to_list(X)))) || X <- lists:seq(1,500)].
    [C:put(riak_object:new(<<"test">>, list_to_binary(integer_to_list(X)) ,<<"98123489123498123498123498123893489712348974123789432789234178942318794213897234178912348791234789412378923417894123789412378941237894123789412387943128794312879123478941237894123789412378941239781243789213487914237891423789142378914234">>)) || X <- lists:seq(1,250)].
    io:format("Start time = ~p ~p\n", [time(), now()]), [ spawn(fun() -> {ok, C2} = riak:local_client(), [{ok,[_]} = C2:mapred_bucket(<<"test">>, [{map, {jsfun, <<"Riak.mapValuesJson">>}, none, false}, {reduce, {jsfun, <<"Riak.reduceSum">>}, none, true}]) || _ <- lists:seq(1,6*15)] , if YY == 1 -> os:cmd("say I am done now"), io:format("End time = ~p ~p\n", [time(), now()]); true -> ok end end) || YY <- lists:seq(1,200)].