sched: walt: Rework the walt irq-work mechanism

It's always going to be suboptimal to report load on
a subset of CPUs to the governor, as the CPUs not in the
subset may have experienced certain types of migrations,
resulting in different frequency requirements. This currently
occurs say when the top-task moves from CPU0 to CPU2, and a
small task moves from CPU4 to CPU0 on the MSM8998 or SDM845.
Since we don't report load for CPU2, the governor assumes
that load on CPU0 has gone away and drops the frequency.
Therefore, when reporting load on behalf of intercluster
migrations, report it for all CPUs.

Also, to reduce the probability of multiple irq-works being
scheduled when there are simultaneous intercluster migrations,
use one irq work for all CPUs. Extend this idea to replace
the existing window-rollover irq-work as well.

Change-Id: I6ca15f13de0b8f91b8a0b1489d75870dc795ad5f
Signed-off-by: Vikram Mulukutla <markivx@codeaurora.org>