Transpose the sched_rq_watermark array

Experimental, needs testers, but so far it works on my machine. This should greatly reduce the number of atomic operations done in update_sched_rq_watermark when switching between processed with different priorities.

One unfortunate side effect is that we can't easily print the idle cpumask anymore in alt_sched_debug. I have removed that feature for now, but it can be implemented again if anyone cares.

This MR conflicts with and makes obsolete !10 (closed).