Support the warpspeed scan implementation on Hopper

The warpspeed scan merged in #6811 is only available on Blackwell+. In order to support Hopper, we have to compile out the work stealing code. We should benchmark whether this renders decoupled lookahead slower than the status quo based on decoupled lookback.

We may also decide to not port the kernel to Hopper, if it turns out to be too much work or diminish it's performance advantage over the old implementation.