Spark always creates secure py4j connections between java and python,
but it also allows users to pass in their own connection.  This restores
the ability for users to pass in an _insecure_ connection, though it
forces them to set 'spark.python.allowInsecurePy4j=true' and still
issues a warning.

Added test cases verifying the failure without the extra configuration,
and verifying things still work with an insecure configuration (in
particular, accumulators, as those were broken with an insecure py4j
gateway before).