Shortcut when switching evaluators: instead of going round the normal
scheduler loop, just cut to the chase and run the thread using the
other evaluator.

This avoids doing stack squeezing each time we switch evaluators,
which is an O(n) operation these days, whereas it used to be O(n) the
first time, and O(1) thereafter if the stack hadn't changed too much.
This is a problem that we should perhaps address separately, but for
now the workaround should provide a speed boost to GHCi on the HEAD.