sched/deadline: Return best-fit CPUs in global CPU DL mgmt for asymmetric CPU capacities

Instead of returning all fitting CPUs for a task, only return those with
the minimal CPU (original) capacity.
This potentially leaves CPUs with higher (original) capacity for more
demanding (i.e. higher dl_density) tasks and so can avoid migrations due
to the fact that there won't be any fitting CPUs free (from DL
perspective) when those tasks get enqueued.

This patch reworks the idea initially presented in:

Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>