Here's a patch, tested minimally by running

    (par-for-each (lambda (x)
                    (monitor
                      (sleep 1)
                      (display "foo\n")))
                  (iota 10))

on a quad-core.  Previously it would print the "foo"s in groups of four
with a second between each group; now it prints them one by one with a
second between each, as should be.