Allow cancelation on async models
Fixes #2212.
Previously, we had `runner.is_busy()`, which for sync models returned True if a
prediction was currently being processed and False otherwise. But for
concurrent models, we can be partially busy.
The cancel method had a fail-fast check to return if we is_busy() returned
False. But with concurrent models, is_busy() has been generalized to mean "is
completely full; cannot accept more work". This meant we could only cancel
predictions if we were currently processing concurrency.max predictions.
As it turns out, this fail-fast check is unnecessary. runner.cancel() checks
that the targeted prediction exists and isn't done already. So we can just drop
this check.
