CRuby first attempts to call `bytes` on the given random object
in order to allow random number implementations to override only
that one method in order to change the random number source. We
were skipping this logic and going straight back to our internal
PRNG logic.

This had two effects:

* SecureRandom and similar classes would not always be the source
  for random numbers they appear to generate, since this callback
  to `bytes` was not happening for long-ranged random numbers.
* The test_ulong_limited_no_rand test from CRuby's test_rand.rb
  failed 50% of the time because we continued to generate actual
  random numbers (1..2) rather than using the non-random
  overridden `bytes` method.